# Devconnect ARG - M2 Yellow Pavilion

## Метаданные

- **Канал:** Ethereum Foundation
- **YouTube:** https://www.youtube.com/watch?v=XL_Nn4oep6M
- **Источник:** https://ekstraktznaniy.ru/video/45115

## Транскрипт

### Segment 1 (00:00 - 05:00) []

to your preference and to your intent. I could go on but but I think those two systems are what's really fascinating here is AI agents also helping you on that execution because that's also a very uh inefficient uh ecosystem — and how would it change like your own life like beyond the million users like the billion users right what like work are you doing right now where something that could not happen today but in the future you're like this is going to be like life-changing for me to happen — for our personal lives — for personal life like beyond like the great UX and the smooth interfaces like right now what's like super annoying and inconvenient to you guys is like going to be changed by what you're working on. I think my life is great right now when I use AI tools but it's very early like what's going to happen is that we are giving like these companies like an immense amount of data like they know everything about us uh and uh essentially the time will come where they need to extract profits from us right it's an early industry like now it's kind of the investment cycle like uh everyone is trying to build capabilities but like the same thing that happened with web one and web two like the time will come where like uh this company need to return value to their shareholders and that's where I think uh things will start getting like worse uh and I think we need to be ready for that moment right like we need to build uh we need to like make progress in like personal AI security uh AI running in like uh tees we'll have a couple of talks about that uh but I think that's the moment where like you could really start seeing that okay my delta of using like my personal agent and connecting to this open market is what's going to make the difference. — So, I'm going to disagree with David because I think that my life is pretty bad when I use AI. Um, I think AI is like actually my like extremely spicy take because I think AI is actually a lot worse than we think. Uh AI is like, you know, it's extremely knowledgeable about certain things and it's very useful to be able to like ask a question to AI and it gives you a pretty good answer, but it's very limited in its capabilities. And like MCP is a great standard, but the thing I hate about MCP is you have to add each MCP server kind of artisally. And like half of them you have to like SSO in. You have to like, you know, connect an account or you have to go get an API key in order for the MCP to work. And you kind of have this like friction every time you want your AI to do something for you that isn't immediately in its capability set. And you end up with like these AIs that kind of pause all the time and be like, well, I can't actually do that for you. if you're trying to get it to do anything non-trivial. And so as you have these open standards, what you'll end up with is uh the ability for your AI tools to have much more power in what they can do for you. And so an example of that being like I internally ask questions to my Gemini or my Claude or my CHBT that are maybe not answerable by content on the open internet. like maybe I want questions where like it's a research paper that's behind a payw wall or like an article that I want it to analyze for me that you know it can't actually access. And so I'm looking forward to a world where AI can automatically add the tools it needs to be better at performing tasks for me and pay the small amounts of money to like just solve problems for me. So my life um well I I've built a bunch of companies and uh managing well max I think 100 people and it was tough. So I'm very excited about the idea of like funding companies with less people and more AIS like as a startuper I'm like feels to me super cool and also like many in different industries. So up to now like the idea just of the idea of I don't know I want to do something in energy I probably need to study six years and it's clear that maybe not giving deep tech contribution but uh entrepreneurs have really the opportunity during their the span of their life to to see more things more industries and that's I think super exciting. Uh maybe something more personal. Um I don't have a driving license, just I'm incredibly myopic because my parents forgot to bring me to the eye doctor until I started bumping into walls and uh and so finally I could have like my own agent bring me around. Uh and uh but it seems a stupid thing, but I'm pretty excited about this. I've been waiting for this. Uh yeah, for me I think what has me probably most excited is just the democratizing nature that this could have right and if that's a developer in uh in in a underserved area or maybe someone who doesn't have access to uh the best practices again of building in this ecosystem to me that's just personally what's I think most exciting is seeing how developers and how um the brightest minds out there can really build on top of that and uh again not only in terms of financial use cases but again arts, science, culture

### Segment 2 (05:00 - 10:00) [5:00]

uh pop culture, mimemetics, memes. Um as we all know, uh and these are really information markets. You know, we're getting into this space here and I think it just gets incredibly complex and yet the AI tools try and help, I think, provide us context as humans in terms of what's going on and it's important to, you know, just be mindful of those gaps. Um but yeah, that that's I think just a similar way I look at it is where uh it may be a show of hands. Uh has anyone here joined the uh ETH uh the ERC community calls over the past couple months? — Yeah, I see a lot of hands here. I think we're up to four calls now. Is that — three? Okay. So, — so you were there. Wow. — Yeah, I was there. I mean, I'm in Singapore, so it's a little late for me, but uh yeah, I'm there. It's also on YouTube. — Take the time. — Yeah. Um but no those are great calls and I think what excites me most and I think for all of us here today uh there's just an incredible amount of ideas on these calls and all the panels today I mean it's a stacked panel uh stacked set of talks today I encourage everyone here to listen um but two parts one there's people here that are speaking about really the silicon level of trust we're at the trustless agent day right trust starts with hardware starts with circuits starts with silicon there's experts in the crowd that'll be speaking later on that very exciting and then all the way up to the user level, right? You know what this means. So, just I encourage everyone to stick here, connect with each other, chat. That what gets me excited. — Awesome. I think we're out of time. Jesse — about to close. We think a couple of minutes marks. — Okay, perfect. Because like there's so much to cover. Like one last part to close would be how do you guys like work together? Where does this all like come together as a closing thought? Let's leave that with that. Where do all the three pillars intersect? I think there is more pillars. I think today is a start like uh we decide to focus on 804 X42 like these are the two protocols that are on top of mind for people. We are making the most progress over the past few months but there will be more and uh we need to continue on this trend. So I hope that the trustless agent community like will be this community that kind of like embraces uh all these people that are building this like uh agentic uh internet. — Yeah. What has happened in the last three months is really meaningful I think in terms of culture and creation of energy and dev engagement. So we need to continue on this with a wider approach. for example the web of agents we need to involve the general AI space not just the web3 space and at the same time while we keep the goal and the dream ambitious that it should be doing the of course the small steps like well shipping on mainet creating an explorer that's I think the challenge of working in parallel on these two levels — and adoption which I think Eric — yeah I think there's no one person in the world who can define and keep in their head all of the standards and have like a deep enough thought process on all of these standards that we're going to need in order to deliver the agentic web. And you're starting to see all a bunch of standards pop up, right? Like I MCP, ADA, A04, X42, each one of these actually even like ACP, there's like no end of these. There's two ACPs, right? And like they each do like a unique useful thing in this emergent economy. And I think that's good, right? Like the I think there's no one group that can actually solve every problem because there's so many problems to solve. And so I think the key thing I try to focus on with X42 is interoperability and being an easy tool to leverage alongside ADA or MCP or 804. And m I think that the to some degree you need to just as you're building these things think about trustless interop and trustless compatibility where these things can complement each other rather than compete because the opportunity is just so big right we have we're talking about the next internet really and that is an opportunity where there's I don't think any one company or person or nation state even is going to be able to define every piece of the infrastructure that's needed for that. — Yeah. And just plus one of these points here. I mean, we're really talking about public blockchains, permissionless systems, public provenence, and with trustless systems, attestability, right? The components of the system that might not be trusted. They can attest to the signal that you feel you may trust, but there's still user choice along your way. And I think in terms of how we can all work together, I think that's what's most exciting is is again it's very much a hardware conversation. It's a utility conversation. It's hardware supply chain conversations. I mean, this goes very deep. The deeper you look into this in terms of where does trust stop or start, you could probably argue it starts literally at the uh at the motherboard, right?

### Segment 3 (10:00 - 15:00) [10:00]

Goes up from there uh all the way through the supply chain up into that user browser on a screen on a mobile device. and the vectors in that space are just incredible, right? So, it's obviously really a team uh you know, family effort here. Um but collectively, I think uh again, it's you know, the outcome here, you know, to Eric's point and everyone else's point here is that uh yeah, there's something you know, potentially uh unique and u empowering here. — I like that you called it a family effort. So, let's close on that note. Thanks so much, guys. Thanks for being here. This was incredible. We not taking questions. No. Yeah, we're done. Thank you. — Thanks, Copa. Thank you. — Uh, is it activated? Hello. Okay, cool. Next up and we have Chi Wong from um Google DeepMind. Uh, he's got it. Yeah. So, Chi is super fun fact. He's been doing programming since he's 7 years old. And for college, he was like building this program to generate some random topics for his studies to choose from, you know, the program that was generated. And this was before AI, um, was a thing. So, uh, he flew all the way from Seattle to be here with us. Um, I think they are still putting together the deck. Um, feel free to like put it up on the screen. And yeah, so he will be talking to us about the frontiers of Asgent AI. He built this framework called Masten. Uh but without further ado, I'll leave it to Chiwan to introduce himself. Thank you. Thank you very much. Thank you very Uh yeah, my name is Ch and uh it's my great pleasure to be here. Uh my background is in AI, so it's exciting to kind of uh share with this part of the world uh and learn from all of you. Uh I'm currently working at Google re Google demand as a research scientist and I'm also founder of multiple open source projects like autogen/2. Uh today I want to take you on a journey through where we have been where we are now and where we're heading with agentic AI. Uh so it all started with uh like when I was seven years old uh my I met my first computer uh it's a Apple 2 uh and it did it had no apps no internet just a blinking cursor on screen and endless possibility. Uh coding in basic at that time felt like magic. Uh I remember thinking like oh this program looks alive. Uh and that illusion planted a dream uh for me to build a AI that can could learn and grow with me. Uh and fast forward to like 30 years later when uh in 2023 uh on a quiet night during a trip to Vancouver that voice came back after 30 years and saying now is the time. So many people were skeptical at that time. Um but I uh couldn't ignore it and uh just decided to drop everything else and to build what I've always imagined together with a community that shared the same vision and that's how autogen was born. Uh the journey of AG2 and autogen it quickly evolved into AG2 which became a foundation for researchers and developers to uh work on the agentic AI space. uh and the journey of AG2 is a good reflection of this wave of agenic AI. uh it um started as a uh multi- aent conversation framework inside uh another open source project called Flammo is for automated machine learning hyperparameter tuning uh I created like six years ago uh and um it's initially was built just for uh helping with our own research uh but it quickly became something bigger uh after spin-off uh within several months it got top training on GitHub uh received uh tops cited by top papers uh and uh got adopted by like tens of thousands of developers. Uh the most fun part is to watch the developers, orchestrate agents like musicians each improvising together uh harmonizing and uh the users of A2 has made new records on many challenging benchmarks like OS world for computer use, SW bench for software engineering and other uh web automation tasks and so on. And these uh progress is not just represent uh the achievements for AG2 but the uh progress for the entire agentic AI research. And the key concepts in AG2 is uh the conversible agent and conversation programming. We want developers to orchestrate multiple uh arbitrary

### Segment 4 (15:00 - 20:00) [15:00]

complex applications in simple ways. Two simple steps. Uh the first step is define agents of different roles and the second step is get them to talk. The key concept here is a conversible agent that unifies multiple different types of backends like language models, non- language model tools and human inputs. And the way you con you build more complex and more powerful agent is to or these primitive agents using relatively simple back ends and apply multiple conversation patterns. Uh we offer multiple uh sequential chats, ny chat, group chats these flexible conversation patterns as building blocks. So you can then compose u very complex multi- aent system and today a2 powers uh many production use cases uh among of many different uh verticals and horizontals uh for example in for agent platform at Google uh for chip design at uh media for recommendation marketing for Walmart and so on. Uh so from uh trading floors to even farm fields uh a2 is changing the agent AI is changing how people work and but so it's not just an experiment it's uh it's a real revolution but we're still very early in that revolution because the if I look around the agent that I dreamed about that can grow with me still doesn't exist but the uh influence of a has reached far beyond these individual use cases and enables uh new research inspires new collaboration, new ecosystems. [snorts] Uh and so the this progress is made in the past three years in the entire industry made me confident that we'll reach the northstar and there's a path forward. Uh to explain that path uh let me walk you through the three key dimensions. Uh the interface, the capability and the architecture. Dimension one interface. Uh today's AI is gaining sound and sight. Uh there are products like Gemini live that can use your camera to see the world around you and then speak to you in real time and um there are even uh prototypes that can see uh what happens on your screen and trigger actions on your side as you use your device. These advances narrow the gap between intent and understanding. But to reach our northstar, we need to do more. Uh for example, the contacts must flow across devices and uh agents must know when to engage uh and when to stand back. Dimension two is about capability. So um a the uh the understanding is just the beginning and the agents uh real capability means agents must act uh on understanding. There are strong coding agents like uh cloud code, gemini, uh codeex agent and so on uh that can integrate into developers workflows and autonomous write uh and modify code but this coding as multip power is still under utilized. Uh agents that can code are supposed to be able to also modify its own software, experiment and learn. And in future agents uh should be able to grow from experience, not just from memory. But to realize this vision, we do need like long-term memory cross- session reflection and iterative improvement. If you think about that uh that's how like entire human society is making progress, getting new knowledge through the scientific process. So science is a very good test um test ground for this agent AI capability. Uh there's a open source project uh called dinario. It's a very new open source project that can help with accelerate the entire scientific process. uh so all the way from the idea generation to writing papers and it's built on top of AG2 uh and uh these type of agents don't replace scientists but they amplify our cur curiosity and uh make it much quicker from hypothesis to insights. And uh there have been like scientists from many different places like uh Cambridge, Harvard, Google and many other institutions across the world. uh and they generated hundreds of papers in different domains uh such as biology, biophysics, uh biomatic uh informatics, chemistry, planetary science, mathematical physics, quantum physics plus cosmology plus machine learning, medicine, neuroscience and mur material science. And there's a recent conference of AI agents for science and the uh one paper fully generated by denario uh is accepted to that conference uh after reviews. Uh so that's is some evidence that uh

### Segment 5 (20:00 - 25:00) [20:00]

the quality of paper but we still need to do more to verify uh their quality in many different disciplines and uh the papers are not the end of it right because no matter how many good research papers are out there they don't add value to you directly until uh you find a way to integrate into your application and someone has been solving that problem the g rank from uh remix AI uh is trying to work on that problem and uh turn the research into a testable uh pull requests in just minutes. It automates the process of discovering the relevant research to your real world application uh do automates the testing and integrate the results. Um so yeah so this uh will um change from just reading about the innovation to actually using it and dimension three is about architecture. Uh in future the breakthroughs will come from uh diverse agents working together. Ag2 orchestrates multi- agent teams and massgen a very new open source project uh push that boundary further by running multiple agents in parallel uh makes each other uh learn from each other and refine result and make improvement. Uh so it's inference time scaling and it can have potential to do expend scaling. So every time you have a new battery agent added to your system, it makes the whole system uh stronger and uh there's active kind of exploration of this new type of uh scaling law. Uh this is uh one example of research uh joint work with my colleague from Google cloud and also collaborators from MIT uh that studies uh like improving uh the improving this scaling behaviors and achieve better performance on challenge benchmarks like last year exams uh and this shows um how this um number of agent when number agents and tokens and inference uh cost scales uh how good the performance improves. So we see that uh it's scales better than alternative approaches but still we haven't reached the ideal scaling curve. So that's uh that requires a lot more research. Uh yeah so u this as I mentioned there are a lot of challenges like consensus interpretability uh and um uh to solve this problem I we do need a lot of help. Uh so please check the project on GitHub. Uh join our discord community and um help us solving this scaling problem together. Thank you very much. Awesome. Uh so while we're transitioning um does anyone have any questions for don't leave yet. — Yeah. — Yes. Do we have a mic? Can you uh grab the chairs now? Yeah. — Oh, thank Yeah. — Uh thanks for sh Okay. Uh thanks for sharing and uh I have a question for your final benchmark. We all know Gemini 3 is amazing but it looks and also we believe that uh with longer inference time more token agent should perform better but it looks like the scaling curve now not even linear. So yeah, what's your expectation or what uh what kind of reason do you infer it could be the like uh — Yeah. Great. Great question. As you see um we don't know yet what is the right dimension about the scaling. What's the right x-axis and the y-axis to expect a very good scaling curve. uh probably the access is not the actual accuracy but some kind of um information about how good the performance is getting improved as we add uh as we increase the inference cost. So that's very active research question we still need to explore and the challenges include uh how you um share the right information share right context with each other how to get consensus um to select the best answer instead of select the wrong answer. Uh and there's also interoperability as different models or agents have very different format. Uh so these are all the different uh hard challenges we want to solve. Uh but we're making quite nice progress in the mastering project and hope we can work on that together. Yes, thank you very much. Thank you. — Okay, I think that's Can we can you ask next question? Okay, I think we h that's about time for uh our current question or our current — speech. So, the team will transition uh and we will have our next speaker once the team helps get the chair down. But

### Segment 6 (25:00 - 30:00) [25:00]

in the meantime, if you look at this QR code, feel free to scan it and submit your questions on Mircat. This is a better way for us to do Q&A going forward. Awesome. — Thank you. — Okay. Awesome. uation. — Okay. And next up we have Nicola Breco from area. Please welcome him on stage. He'll be talking about what if agents could create protocols on the fly and welcome. Thank you. — Hey everyone, thank you for organizing this event. Uh this is not my slide deck but I would be down to present this one. Uh oh. Well, I'm just going to say a few things. I think that um this is like a milestone event. The way I see it, we start to see like a few standards for interperability and this is like one of the first conferences to talk about this in a very u decentralized and open network way. Okay, I'm going to say a few things about myself. I am Nicola. I led the cryptography R&D teams at Protocol Labs and then at FODS and I'm collaborating uh recently with Arya which is a UK R&D funding organization uh that funds um uh R&D moonshots and one thing that has been always uh something that I cared about was um can we bring cryptography outside of the academic circle outside of uh the cryptocurrency industry and in the hands of everyone and so today we're going to talk about what if agents could write protocol on the fly in such a way that everyone then is able to use uh these uh advanced cryptography tools. So let's say that I want to have this interaction and uh it's uh hopefully very soon thanks to all these protocols and I want to book a meeting with Alex for tomorrow my collaborator and I tell my agent uh when are you available tomorrow and then for whatever reason the agent says sorry cannot disclose the information of the of uh of this other party. Now this is an interaction which is a fictitious. I'm sure they he can share these details but maybe we might end up in interactions where the other agent cannot move forward in the conversation because they would be leaking information about their principle. And so how do we move forward? How do we make sure that this interaction happens? And so one way to do it is by having centralized mediators maybe like a calendar app or maybe to some extent stock exchanges apps like Tinder. They all act as mediators for our private secrets. Another way that I would think as a cryptographer, I would hire the best engineers to implement a cryptographic circuit so that they could have this conversation in a secure way without a trusted mediator. But this is very expensive. Like we cannot reasonably do this in our day-to-day life or our day-to-day interactions. But AI can write programs on the spot. Maybe today they cannot do them well. And just I just want to say something important. This talk is not about what we could do today but what I would like that we would be able to do in the far in foreseeable future. So the question comes natural which is can AI agents also write protocols and maybe the interaction is going to be like hey how about you're not you don't want to share information with me how about we run this circuit that I just designed for you to interact with me. Deal. This interaction would be fantastic. And how can we get there? First of all, agents need to be able to reason about the right primitives that they want to use. They need to be able to understand the interaction. And not only they must be able to implement it and on the other on the receiving side verify it. And I think this could open up to a whole new range of cryptographic interactions. And if AI agents could write protocols then there is this uh book um at school you told that with MPC two satellites can try to not uh collide with each other by running MPC. Well now they can do that and many other things and maybe uh if you want to sell your extra storage on your computer maybe you can ask your agent figure out a way to sell it and they will figure out a proof of storage that they can use and convince the other party that that's good enough. And you can do a lot more. You can do uh zero knowledge proofs. You can write smart contracts. You can use uh trust the hardware. And there's a lot more. And I want to call all of this generative crypto. You can think of this as autonomous protocol design where the agent is able to independently come up with the right thing to do. And I want to um say like uh a few there

### Segment 7 (30:00 - 35:00) [30:00]

are a few things that I think are very important uh and why this is and why we should work uh towards this. So the first one I think uh is straightforward safer interactions uh agents if they want to run um information uh interactions with private data they're forced to leak private information or rely on a third party or they're not able to interact. The next one is what I like to call AI advantage. Um, so there is something that AI can do that is uh that humans cannot do and one thing is supercognition. So humans cannot just uh think on the spot how do you run the uh what is the right protocol to implement and even if they could because they're great they cannot run computers with name brains and but AI can and AI can do a lot more things beyond uh coming up with the right protocols and the second thing that another example of AI advantage is that if two people try to make a deal and then and they must exchange information private information to make a deal and then the deal doesn't go through then they leave then leave the meeting room and go back to their companies with secret information. Like if we want to make sure that if they leave the deal, they don't have the information anymore. I cannot think of anything that is not illegal to make sure that people forget information. But what we can do if the agents running a te agents can forget information in the if the deal doesn't go through and this kind of private interaction don't exist uh today. And then finally a few uh researchers are writing great academic article and blog post where they say that we might reach this coasian singularity meaning that the uh transaction play an important cost in our society they shape organization and our economic structure and lower transaction cost reduce the threshold of what's worth doing and transaction cost is not what we mean uh in the cryptocurrency industry transaction cost is the cost of discovery execution negotiation enforcement and so on and agents can lower this transaction cost because they can interact in ways they couldn't before. They are not bounded by limitation such like uh fatigue, bias, logistics and so on. And my the case that I'm trying to make is agents are lowering down the cost. We already can see like search of for agents it's really good and they help us finding better products. But if these agents could also tap into this uh new trust uh infrastructure through generative cryptography then they're able to uh further lower down this cost of trust. So they lower this trust overhead. And the last item is something that's very close to my heart which is I do think that giving cryptography to everyone enables this uh ultimate cipher punk dream. And you can think of the web is a set of few things that made it trustworthy. encryption, digital signatures, sandboxing and these small trust building blocks gave us fintech, secure messaging, e-commerce, digital cloud, um digital ID. Essentially, these small trust primitives that we all have in the browser allowed us to participate in this digital society and um there is more to encryption and digital signature. There's like a lot of cryptography that is application specific. There is programmable cryptography that you can write. But how many of you can write uh secure cryptographic circuits like I don't know zk circuits one that's awesome you see it's hard if we can make it simple that so that agents could write the cognitive load is so high so if we could make it simple so that the agents can write code on their own then we could have what I call cryptography for everyone and it's not just cryptography this is very important because some interactions you cannot do them with cryptography you trusted hardware. It's like a protocol designer for everyone. It's like an engineer in your browser or an engineer that any of your agents can tap into in order to have these sort of protocol interactions with each other and uh and so in a way it's a way to scale trust uh to for everyone. And so where are we at? So can AI agents reason about security? Can they compose protocols? Can they write them? Can they write security proofs? Can they come up with new protocols? There is like a large uh there is a large spectrum. They could be a protocol designer or an AI scientist. Where are we? I don't know six, seven. Um so where are we at? So I would say that most cryptographers are already using copilots as like AI agents to write um as they write papers they ask the AI, hey do you think this is correct? The second thing is I believe I hope that someone is going to go out of this room today and go build it. We can have agents that use application specific cryptography or TE's through uh

### Segment 8 (35:00 - 40:00) [35:00]

something like uh MCP and so you can you can make a cryptography library hooked into uh an agent. So this you can do but I think having AI agents that can come up with cryptographic protocols doesn't have to write new code they just need to come up with the right thing to do for the policy of the user this we cannot do today and that's why at area we are looking at this problem specifically and then there's this even harder problem which is do can we have AI agents can AI solve uh new problems or write complete new protocols that we didn't know we could so pick your favorite cryptographer and think what do cryptography ers do. They identify security goals, they design security protocols and so they design the protocol, they prove them secure and they implement them. And the beauty of um this way to think of this is that we can layer these things um uh in a in a very uh thin way which is we could have a reasoning engine that can go from user definition to a protocol that could be implemented with a trusted party. You can have a cryptographic engine that takes this protocol and turns it into like a cryptographic protocol. So no trusted parties anymore. And then you can take the spec that the previous agent produced to generate uh working code. And this can all be different agents. It could be that there is one agent that does everything. And but I think focusing on the first one will be uh the most impactful one. And you know we could break down the problem in multiple parts. I think the agent needs to be able to require gather requirements. They need reason and negotiate and these all come with a bunch of these different questions. So at area we're designing this new program and there's going to be uh 50 million pounds in R&D funding for the intersection of AI security and cyber physical systems. We didn't talk about cyber physical systems but it's a core part of our pro of our program and we announced the thesis two days ago. So I would love to hear uh all of your feedback and the way we want to do this is not necessarily by implementing everything by creating benchmarks and competition where people can go and compete and reward uh the winners and uh we are still designing uh the program and we would love to hear your um your opinion, your feedback, your interest and uh there is a lot of safety concerns that are very important that we need to make sure that we reason as we do this. So we're also looking for people that can help us in this way. To give an example, if AI agents can write security protocols with each other, maybe we don't know what's going on around us. So maybe we need to find guard rails for that. Anyway, thanks a lot. I leave you with this comic. And I am not Nicola on X or Telegram. And thanks a lot everyone for having me here. — Thank you. Thank you, Nicola. Um, does anyone have a quick question or two? Maybe feel free to find Nicola as well offstage. I think feel free to like come and hang out with him. Um, so that will be that was really cool. Uh, a lot of cyber physical I think I haven't seen too much on that yet. But next up we have Quintis from Flashbots. He will be talking to us about trust through silicon. Welcome Quintis. — Thanks. AI at the same abstraction level as you guys. Um, I wanted to start the talk today with a bit of an odd quote or a question. Can you keep secrets when an adversary can eaves drop on your brain? Um, as I go through the talk, it'll make more sense and the origin of the quote is actually kind of important. Um, but to understand the importance of this question, we can ask a simpler question first. What if your brain couldn't keep secrets? I think you'd be kind of sad. Uh firstly, you wouldn't be able to hold any passwords. Um and you would go to a market and probably pay the highest price you possibly could at every time because they'd know how much you'd be willing to pay, but your friends wouldn't confide in you. You wouldn't be able to be a lawyer or a doctor or a therapist or work at a company uh that had NDA requirements. Um and as people who are designing brains uh in this audience both at the agent level and at the emergent level uh I think we need to think of secrecy as an engineering problem. The ability to keep secrets is an engineering problem because it enables coordination. Uh and one level of where we want to get to is like at least getting to where people are where we can you if we want to keep

### Segment 9 (40:00 - 45:00) [40:00]

secrets. But uh if we think of it as an engineering problem, maybe like uh Nicolola said, we can use an AI advantage and go beyond what people can do and make a really strong statements about how information is being used to push coordination beyond what we can do as people. Um and we already have tools for this and we're already using these tools, right? And so like uh more than a decade ago, we realized our relationship with silicon had this problem. Um silicon uh doesn't tell us much about who's eavesdropping on our interactions. We'd send data to some chip, get a response, especially if the chips in the cloud. We have no idea how our data is being handled uh or whether the response is uh you know coming from the same software that we expected. Uh and as a response to this, the idea of trusted execution environments came along. uh and I won't explain it in too much detail but the basic idea is the chip is designed with some security properties and a embedded cryptographic key so that a remote user can interact with the with a program and have some confidence that uh no one's eavesdropping on the execution uh and that the execution of the program is being done correctly um and this is already being used for like confidential inference uh protecting model weights and these kinds of things but I was talking about sort of what we can do uh beyond on the standard kind of keeping of secrets. Uh what's interesting for coordination? Uh so Nicolola already kind of gave this example. I'll just fly through it. Imagine we have two uh manufacturing businesses. Um and each of them has a sort of proprietary process uh that allows them to make the money they do. It's possible that if they combine these processes, they could come up with some even better process that was better than the sum of its parts. But at the same time there's risk that if they share this process with their counterparty uh and the deal doesn't go through they've ruined the business opportunity. Um so the solution that I propose here is that uh each business uh trains an agent on their confidential information and submits it into a black box. And the black box gives us a property that the only output uh of the computation will be deal or no deal. Uh and we can proceed based on that. And this uh removes the dilemma that the uh business owners were in uh where the risk is now gone. there's no risk that the interaction will reveal their proprietary information and of course uh I'll say that we can implement this with TEEES it's possible today uh and there's a really interesting paper about this so this is a bit of an abstract example but I promise uh that this is something that sort of a kind of dynamic that appears in everyday life quite often and definitely in my work in ME we see this kind of dynamic quite frequently uh but I said I was going to talk about silicon and that's also very high level um all of what I've been talking about so far is assuming that we can keep secrets. Uh but now maybe it's time to talk about where that quote came from. In 2003, this paper came out with the initial quote that I gave you as the opening line. And um you what is the motivation of this paper? Well, basically this graph communicates it. This is the electromagnetic radiation of an ARM processor doing division on different values. And as you can see, anyone who can uh observe the radiation coming off the chip can learn the computations, the secrets that are being um computed inside this processor, inside this brain. Um and okay, division is kind of boring. Uh but there's actually many different papers that have demonstrated that attacks that use this kind of information, side channel information, can be used to extract model weights, uh to reverse engineer uh model inputs. This uh this screenshot over here is um shows how by purely observing the encrypted memory of a TE uh poorly encrypted uh the researchers were able to reverse engineer images that they had not seen before or that their model uh very accurately. Um so eavesdropping is real, right? But who are the adversaries who can do this? Um the most obvious adversary is the data center. That's where a lot of the compute uh happens today. And um I think in the short run, the data center does give us a lot of protections over just nothing over trust me bro. Um but at the same time if we're trying to build this new economy where we have multiple agents interacting um and we think that that's where a lot of the human or a lot of the economic coordination is going uh then having this sort of corporate governmental entity being able to eaves drop on every agent's thoughts on every interaction probably isn't good for the sort of the power balances of the world but also just for the economic flourishing uh when we think of secrecy as an enabler for coordination. Uh but if that's too like abstract cippher punk for you, I

### Segment 10 (45:00 - 50:00) [45:00]

think there's a more realistic kind of answer. Uh which is the cyber physical world as Nicola put it. Uh we plan to have agents and models uh embedded in the physical world around us exposed to many different physical adversaries um in our self-driving cars, in robots, in uh BCI, in surveillance cameras. We want to have confidence uh that the data from these um different uh machines is being used in a way uh that we approve of or at least we know how it's being used. Um and from sensors we also want to know that the sensor data is like not being generated uh by a model. So that was the physical side of things. The another important and a very difficult to address challenge is that most of the hardware we rely on today has this closed inner kernel right the firmware the hardware is completely opaque to us. Uh and you know in the AI world you know very well that there are only a very few companies that provide most of the silicon we rely on and that means that they basically have this uh backdoor into this uh agentic economy that you guys are trying to build. Um and so in combination between having to trust the data center and um chip manufacturer and the entities along the supply chain, what we call secure hardware today is basically just a wrapper around the good favor of large corporations uh who can eavesdrop um on the brains of the brains that we're building. Now this sounds very gloomy uh but these problems are solvable and you'll just have to take my word for it now. I don't have time to go through uh all of that in this talk. Um but the these problems also raise a next set of problems that I think this audience is well equipped to think through. How can agents or people reason uh about secrecy in this heterogeneous world where we're relying on uh silicon and maybe other techniques to give us this kind of eavesdropping uh protection when there are different supply chains, different actors that we trust at different levels uh that we know more or less about. Um, we have different verification standards. One being just the manufacturer saying, "This is my hardware and I promise it's fine. " And another going through very arduous testing with different labs, sampling the chips, checking for back doors and these kinds of things. Um, and different physical security mechanisms. uh the fact that most of the hardware world and a lot of the software around it is closed makes it very hard to compare um the security mechanisms implemented in these ships across each other and just at the software level as well software level as well. uh at the same time the existing certification bodies like common criteria are uh not very good um and don't tell us necessarily how uh what the actual level of security is and so it's a very complex world to reason through and I don't have great answers uh but if we want a world where uh agents are freely interacting with all the different uh embodiment of intelligence around us around them um we're going to have to think through some of these problems and I don't think a good solution is only trust the hardware manufactured by Nvidia and Intel or an AMD. Uh we want to have a more diverse uh world. Uh some basic solutions are like uh having public bounties so that anyone who can break one of these chips can claim a bounty to give us some economic assurances that are transportable across these different uh settings. Um I'm running out of time so I won't go into it too much. The one cover I wanted to give here is that FHE doesn't remove the problem necessarily. it can improve it somewhat. Uh but you still, you know, there's no such thing as fully homorphic decryption. You still need to re rely on hardware for some of the collusive attacks, etc. Um and that's basically the crux of my talk. If you're interested in what it means to uh improve uh security of silicon in the different ways that I mention here, uh you can look at what I'm doing at the trustee initiative with my collaborators. Um and thank you for your time. — Thank you so much, Quintis from Flashbots. Um, if you have any questions, feel free to catch them after uh under right there over there. Um, and uh next up we have Molly from IPFS. Uh, she'll be talking to us about Deepen for AI. Let's welcome up for Molly. — Thank you. Hello everyone. Microphone working? We all good? I'm so excited to talk to you all today. Um, I'm Molly McKinley. I'm the CEO of Phil Oz, which is a core R&D team in the Filecoin ecosystem. I also was back in the day in I think 20189 the project lead for IPFS. I've gotten

### Segment 11 (50:00 - 55:00) [50:00]

the real honor to pull these two ecosystems um even closer together. they're all built on the same underlying stack. Um, and so now I'll be talking about first how AI is creating massive scale demand for DPIN, but also the tools that we are building in Falcoin and IPFS land to support this huge new demand for decentralized storage. Um, okay. So, if you'd like to follow along, there's a QR code so you can scan and skip right ahead to whichever slide suits you best. Um, but looking forward to get started. Oh, sorry, not fast enough. Um, all right. So, I'm sure we all know this, but agents are consuming massive amounts of resources more than ever before from storage to compute, data, payments, and beyond. Um, according to projections, the AI compute market's going to more than 4x in the next 5 years. That's probably an underestimate. I think it's going to go even bigger than that. Um, we're in a crazy era where the demand and scale of resources that we need to create as an ecosystem, and by ecosystem here, I mean the AI ecosystem, not the web 3 ecosystem, um, is scaling incredibly quickly. And when I'm talking about agents here, I'm not really talking about what agents can do. I'm need to operate in their world, what services they need to depend upon in order to get their work done. Um, autonomous agents need robust and self-s sovereign compute, data, storage, and other services that don't have fragile, permissionless, and opaque layers of centralization where they can't control their own workloads. Running trustless agents on centralized stacks are going to fail the core properties that trustless agents require. They need ownership. They need pro public verifiability. And they need permissionlessness to operate in a fully decentralized onchain layer. Growing agent demand is going to create truly massive demand for infrastructure. And the question is on what stack is that infrastructure going to scale. I posit that centralized infrastructure is the wrong choice for permissionless agents operating in web 3 and beyond. They need continuous machine-to-achine global and trust minimized compute, storage, and cloud services to work with their scaling workloads. The web two centralized clouds, I know I worked at Google for a long time, were not built for this type of demand and for the open, trustless operation that we're going to need when working with onchain agents. Deepen on the other hand decentralized physical infrastructure networks were designed for exactly this sort of use case. They are built for open markets and they connect global idle resources so that they can scale in a much more horizontal open way than centralized infrastructure that needs to be controlled and audited by one single company and provider. They're also more robust. We all well at least everyone in my ecosystem was very aware of the Cloudflare outage on Tuesday. took down a whole ton of websites um including a lot of people on uh Twitter and other places. Uh and so we need an infrastructure stack that AI agents can use that is robust and resilient and doesn't go down when a single provider has an outage. With deep pins and agents, we can create a flywheel for cheaper, more resilient, and more trustworthy cloud services versus these legacy centralized cloud systems. So in order to build trustless agents, we need core primitives that the decentralized web has been building for the last decade. We need data set provenence with content addressing and chain of custody over the time from data creation to usage and uh IO with models. We need durable memory for state so that we can audit and make proofs about it over time. We need auditable logs like the um 8004 verification stack so that we can audit uh agents at every step of their path. We need verifiable compute so that we know that auditable inputs go through auditable compute systems and can become trustworthy outputs from models that are maybe operating in autonomous environments without human oversight and input. And we need payments, storage, and all of this operating onchain with zero human intermediaries or points of failure. The centralized clouds just can't meet these demands. That's not how they operate. That's not what they're designed for. But deepen networks offer proofs, not promises. So, I'd like to share with you all uh a new uh storage layer inside Filecoin that we just introduced on Tuesday, which is called the Filecoin onchain cloud. It's a verifiable trustless stack for resilient autonomous agents. Uh it's and it provides the primitives that they need. It has snark verified storage, payerbite fast retrieval, uh IPFS layers on top so that you can content address everything in and out. You have fully provable decentralized

### Segment 12 (55:00 - 60:00) [55:00]

storage so that you can work with truly trustless agents in an on-chain environment. This is perfect for verifiable agent memory, logs, data sets, embeddings, and a lot more. So, the onchain cloud is able to turn any service from indexing to model inference, search and retrieval into a provable function with metered economics. We also on top of the Filecoin onchain cloud built Filecoin pin which is a developer suite with a CLI tooling library and even a GitHub action for publishing pipelines running on top of Filecoin with IPFS content addressing. It's compatible with all of your existing IPFS tooling, Kubo, IPFS gateways, whatever pipeline or ecosystem in the IPFS space you prefer and support. Um, and here you can see our IPFS pin web app demo that you can use right now. If you scan this QR code, you can literally take a photo of me and store it on Filecoin. It'll be your first decentralized Well, maybe some of you have already stored data on Filecoin. Awesome. But, uh, my favorite thing to do is literally take selfies while I'm on stage, store it on Filecoin within just a minute, and, uh, off to the races with decentralized storage that is truly permissionless and robust. Um, so feel free to go ahead and try it out and hopefully I'll see a couple fun images if you tweet them at me. The you can send me the IPFS gateway link so that we can see all of the amazing things we're doing. Um, so you can start using the Filecoin pin service on test net right now. Um, be among the first to also claim your free storage on mainet specifically for 8004 builders this coming January when both 8004 and the Filecoin onchain cloud go to mainet production readiness. So get ready, get excited. We've been really pumped to be building with the 8004 community for uh the past couple of months. Um you can also we've written a couple of getting started docs guides. Check out this Filecoin pin demo video on YouTube um specific to running 8004 agents on Falcoin onchain cloud. Um, and you can also we have an MCP server um that allows you to interact with the Filecoin onchain cloud from your agent directly and do all of your storage and payments and everything uh trustlessly uh in that environment. So I will pause a second so you can grab that QR code as well. Um so why does this matter? What is what are people going to do with this massive scale decentralized storage? They're going to put truly massive amounts of AI training data on it at scale trustlessly so that we don't only get the verifiable outputs of small models, but we get all of the robotics data that Bit Robot is collecting, which is a deep pin network building on Salana using Falcoin for storage, which is going to be generating absolutely massive amounts of diverse robotics data about embodied AI training and workloads. Um, this is the missing agreement ingredient for general purpose physical AI. If we want robots interacting in all sorts of uh creative environments and a whole diversity of robot types, we need the data that is going to power that the same way that we've been doing it for autonomous cars for the past over a decade. building up that data set for robotics is going to be critical for this to happen safely and for us to actually be able to have these uh you know amazing humanoid robots that are getting built be able to interoperate with humans um in real world environments. So over the coming years uh I think AI pin networks are going to be a massive huge collector of pabytes of AI training corpora and filecoin is unlocking our exabytes of storage capacity to support these verifiable data sets um in the open onchain cloud instead of locked in proprietary silos and centralized clouds. So we all know that this scale is truly massive. Um I think over the coming couple years the primary consumers of our distributed decentralized infrastructure is actually going to become agents themselves not humans training agents or interacting with agents. Um these agents are going to be the ones who are utilizing compute who are paying for services who are interacting with each other who are purchasing additional data for their own evolution and growth. And if we want to build trust with them, we need to make sure that we support them with the infrastructure that allows them to use it in a self-s sovereign way. Um, and I think that will fuel a trustless agent Cambrian explosion. Um, that we can all benefit from. All right. So, if you take one thing away from this talk, I hope you agree that trustless agents require trustless services. Agents are only going to be as trust trustworthy as the services that they consume. And agent economies need open, verifiable, self-s sovereign infrastructure that builds trust at

### Segment 13 (60:00 - 65:00) [1:00:00]

scale. At Filecoin, we are deeply, deeply motivated and focused on building fully open, programmable, verifiable cloud services that can support that massive demand to come. So, please join us in building f fully trustless agents on fully trustless infrastructure. Learn more at falcoin. cloud cloud or get in touch with me on Twitter, Farcastaster or anywhere else on the internet. Super excited to build with all of you. Let's make a really beautiful internet that we can all benefit from. Thank you. Thank you so much, Molly. That makes a lot of sense for verifiable data sets for bringing a trust us agent um future. So, next up we have Shafu, who is from Merit Systems. He is doing X42 scan. Let's go. All right. — Hey guys, I'm Shafu from Merit Systems and today I want to talk about X42 scan. So X42 scan, if I figure out the clicker, this one. Okay, X42 scan is the discovery layer for X42. It's something like D5 Lama for DI or Ether Scan. It's like a block explorer for the whole X42 ecosystem. So when we started looking at X42 a couple of weeks ago and maybe some context here, X42 is a very new thing. Uh Eric and the guys at Coinbase released the spec in May and when we started looking into it like 8 weeks ago, what we saw were around 50k transactions. It was very hard to find all the X42 resources. servers that expose X42. And there was also no like one analytics or statistics page. Uh there was one dune dashboard but that was pretty much it. So this is why we built uh X42 scan. We released that in October and it basically lists every X42 transaction that is out there. We list X42 resources. We list uh you can play around with the resources as well. We list all the X42 facilitators and it's one it's the X42 resources. I think this could be the beginning of what people like to call um agentic commerce. It's basically an agent that roams around the internet. You hit free APIs, but you don't get stuck anymore by pay walls. Basically, it's like your agent needs to access a paid article on the New York Times. it goes there. It just pays a couple of cents and gets that data back. Um, so very excited about this. What you see right here is how the actual chat interface looks like very much like something like chat GPT, but in this case it actually cost an X42 resources resource pays like seven cents and I think this is pretty cool. Okay, so one big problem we have with X42 is something that I like to call the reputation problem. Anyone can create X42 resources, which is awesome. But if anyone can create them, you will have some resources that are not that good, right? They just don't work. Their uptime is bad. So the solution to that is give every resource a reputation score, right? And because we at X42 scan we have a lot of data uh that we collect. We have uptime, we have number of transactions, we have volume, we have latency. So we use that to create one aggregated score for every resource and basically give you like a health status. Where this is very interesting is again back to the agentic commerce part. Your agent has its own wallet. it's going to spend your money and you want to make sure that it's actually going to use resources that are good, right? Because you don't want to pay for something and then don't get back anything. We also have some more detailed observability pages. This is very interesting for developers that want to debug their resources. We have the same thing for facilitators as well. So yeah, we launched X42 scan. We got a lot of eyes on it. uh very fast. Most of this was X42 meme coin mania. Um the attention was good, but I think it's time now to actually build some really good useful X42 resources. Um the question I like to ask myself is what's possible now that wasn't possible for before X42 and we need to build that. I'm proud of this. X42 scan is completely open source. The front end is open source. The indexer is open source.

### Segment 14 (65:00 - 70:00) [1:05:00]

The sync the syncer is open source. Even the proxy is open source. Just imagine if etherscan was open source. I love etherscan, but if it were open source, the level of innovation would be completely different. So yeah, come build with us. We have the screenshot is a bit old now. We have over 30 contributors at this point. Uh I think only seven out of these are from the merit systems team. Um, if you're interested in X42, uh, come build X42 scan. So, yeah, maybe lastly, I want to talk about the vision here. Uh, I'm personally very excited about two things in X42. Uh, the first thing is, in my opinion, X42 has the ability to kill ads on the internet. I don't like ads very much. Um, so that's very exciting. Just imagine instead of getting YouTube Premier or something, you just pay a couple of cents, you don't see any ads. Same time with like the New York Times example. Um, the second one is agentic commerce. X42 is perfect for that, right? Agent goes around the internet, hits a payw wall, just pays for it and get back gets back an answer. Um, the screenshot here is something that A16Z posted a while ago where they were basically, hey, the potential market for X42 is the whole internet. So, we're talking about trillions, big numbers. Uh, very exciting stuff. So, yeah, what's next? Um, V2 is finalized. X42, uh, V2 is finalized, which introduces some very interesting features. Uh, shout out to Coinbase for doing it in the open. This is in the Coinbase X42 repo. And, uh, and again, it's very early, so you can contribute to a lot of things. You can do reputation, privacy, indexing. How do you discover things? Even simple things like documentation tutorials will be very useful. So yeah, visit x42can. com if you're interested in x42. Um, come build with us. Let's build. Let's grow the pie. Let's make x42. Let's 100x bigger than it is today. Thank you very much. — Thank you, Shafu. So basically, if you know, X42 was how agents were going to pay each other, you'll still need to find like a way for endpoints that are worth paying, that's what part of the value prop of X42 scan. Uh, does the audience have any questions for Shafu? If you're too shy to ask, oh, — go for it. — Hey, first off, Shafua, thank you for X42 scan. Amazing. Um, woo. Yeah. — Thank you. Thank Thanks. — And the main question I have is about the uh name spacing of the services. So like what I'm going to end up doing for instance is having tons of microervices and each of them has their own reputation and scoring but also we need to aggregate the kind of like the provider underneath as well. So like is that on your V2 road map? — No, it's a very good question. What we currently do is we not only give a resource a reputation score but the origin as well. — So maybe that's — Yeah, I think that'll help. The main thing is just in case the origin isn't declaring itself, how are we going to handle that where people are just like making tons of little tiny services, but they're wrapping other people's services. — So yeah, you would need like an abstraction in the middle where it's not only resource origin, but something Yeah. in between those. — Got it. All right. — Thank you. — Awesome. Uh I think we are running out of time for questions but feel free to come to chat with Shafu in person right over there by the stage. Um thank you Shafu and yeah next up we have open med which is the open source 804 facilitator for X42. So this is kind of like maybe like the first and you know the real bridge between ERC 80004 and X42. and Cody from Bass and David from EFDAI will be presenting to you the latest demo here and the stage is yours with the clicker if you need time to set up the presentation. You can ignore Can we get this lights out? — Okay. — Hi. Yeah. Um, this is a work um I care about a lot. So, essentially like it's bridging 804 and X42 is the first of I think a long series of work we need to do. So, I'm excited to like introduce uh this and the work we've done with Kodi. Yep. Just use the computer. [snorts] Yeah. I want to start like by saying like 804 from my perspective like we

### Segment 15 (70:00 - 75:00) [1:10:00]

have the onchain registry it is also a data structure so it is actually being used uh beyond uh ethereum there is other chains that are interesting of implementing this it could be used in ofchain systems and this is important for today's talk uh it's also a vision right it's it started with a meme of uh trustless agents and that's how like this community got together uh but now we need to work towards this vision and this will mean like going beyond the actual protocol that hits on Ethereum build integration to all these other protocols of the web of agents that we talked about. So what is this web of agents right like so we uh we have an app layer there is a web services AI services there's going to be robots uh multi- aent systems we'll see some examples uh later today games uh then uh you need payments right so there is like the paper call with x42 there's going to be subscriptions there is also going to be no payments right some of these multi- aent system will be just like instantiated and like there is some different type of value transfer and uh a 2004 registry can be like the settlement layer that is like uh below and supporting all this uh providing identity and providing reputation and it's especially important in connection with X42 right like this is kind of like the trust minimized uh settlement uh part of the stack so that's what we are focusing on and we built this uh initial uh extension to X42 uh which natively supports some 804 uh functions. The first one is about identity. So essentially uh there is the 804 identity registry. We want like X42 services to like actually use Ethereum uh like a decentralized like um network for discovery. So we have a register extension where like if you have um service that uses X42 now you can use the facilitator not just for settlement of uh payments but also to help you register on 804 and in the future we also need to build discovery services like decentralized discovery is actually like a big um uh opportunity uh for X42 I think uh because as payments start flowing through this system there is centralization vectors the same we've seen like in other infrastructure. So we want to like decentralize that. The second part is reputation like 2004 also has the reputation registry uh and uh the facilitator can help uh the client submit feedback uh to the services that they use and then in the future we can use this to fetch reputation and to provide like even more information to like um the x42 network payment network. I'll pass it on to Cody to talk about um uh the facilitator. — Uh yep. So uh when I think about the design principles for this facilitator, uh there are a few things I was thinking about. Uh so number one, uh I want there's like minimal setup for the service agent because the beauty of X42 is that it's really easy to wrap the API with X42. And then secondly, I want the facilitator to do most of the work. So for example, these days facilitators all sponsor gas. Uh so we're using 7702 plus a de delegation contract to do this. And then I wanted to tie very smoothly with the X42 flow. Uh then when I reached out to the X42 team, they mentioned there's a X42 V2 that has the extension features which I was using it here as well. And the first part is the register flow. Uh basically the facilitator is going to expose the register endpoint for registration and then the server service agent owner is going to send the 7702 off for delegation along with the token URI and metadata uh which is the data required for registering agent on uh 80004 and then for the feedback flow if you're familiar with 804 uh when doing a feedback you request the service agent to uh sign a feedback off for the corresponding client. So in this case uh after X42 settlement happens the this facilitator is going to generate a feedback O by calling the server to sign the feedback hash. So in this case the server one needs to expose a signed feedback hash endpoint so that the facilitator can ask the server uh to give the permission for the client to submit the feedback and then later on the client calls facilitator's feedback uh endpoint uh also via 772 to submit the feedback on chain and as I mentioned earlier the agent setup here is pretty straightforward uh it firstly needs to make a

### Segment 16 (75:00 - 80:00) [1:15:00]

registration endpoint uh call to the facilitator to register itself uh using 7702 and then later on for the feedback uh as you can see here uh I'm using S42v2 version which has this extension feature which makes it really easy to pass uh variables to let the facilitator to extract the information. So for example, the agent can set up uh some fields here. Uh they can specify if they want the feedback to be enabled, what's their endpoint to be used to sign the off and also their agent ID that's already registered on chain so that the facilitator can extract this information really quickly. Um here I'm going to give a uh quick demo of how it runs. Uh so uh hopefully it's clear for people to see. Um let me so uh here I have like a service so here I have a service agent that's basically the same as I mentioned in the slides. uh it has a weather API that accepts uh 0. 1 cent uh USDC uh via x42 and then it also has this extension feature where uh where it tells the facilitator uh if it wants the feedback to be enabled what's the endpoint to use to sign the O and also the registered agent ID uh on 804 and then on the client side uh the stuff it does is pretty simple Uh this method just fetches the resource by calling the weather API and also it has a submit feedback function which uh makes a call to the facilitator to submit the feedback. Um and I'm going to open up my terminal uh here. Uh I'm doing it locally so that we don't have to deal with the bad internet here. Uh I'm setting up the facilitator at the top. Uh so this facility has all these methods available. So verify settle which is the standard X42 features and also other features like uh feedback and also register uh that's needed for 804 and then uh here I'm setting up uh my server on the left uh on the bottom left hand side and then uh left right hand side I have the client. So I'm going to start making the request to get the resources. Um so when it's getting the resources you can see on the facilitator side uh it's generating a feedback off by calling uh the uh sign feedback off method on the server side. So the a feedback off has already been generated here. And then nextly once the client sees the resource and then when they want to rate the agent now they can make a call uh to submit the feedback uh via the facilitator on chain. So here is subling a score of 80 on this agent and as you can see uh very quickly the facilitator uh pick up this request uh and then submit this uh feedback on chain. So if we just go on chain uh on the base scan here uh we could see that like this feedback has been submitted. Uh yeah. Uh that's pretty much it for this flow. Um and yeah and this is like the overall flow. Uh basically to summarize it's very easy for the server to set up the things needed for 804. You just need a reg you just need to call the register endpoint on the facilitator and also need to expose a sign feedback hash and no gas required for the client nor the for the server. And we're also aiming to have both V1 and V2 compat compatibility for the facilitator. And uh as you can see from the demo, it's uh the main functionality is already there. We'll be live on Saporia soon and it will be open source. So please stay tuned. Uh that's it from us. Yeah. — Okay. So I think we don't have much time for a question but feel free to chat with the speakers over there. Um and next up we have uh Rodrigo from Edge and Node. He'll be talking to us about how Emperson makes agent payments verifiable and controllable. And essentially you know how do you verify agents have payment and payments that are verifiable

### Segment 17 (80:00 - 85:00) [1:20:00]

and that's policy bound. You know think about spending limits approvals audibility for the agents budget. um and he'll be talking about that because it will be really important for enterprises to adopt agent payments. Let's welcome Rodrigo. — Awesome. Here's the clicker. — All right. Yes. Almost. All right. Uh this is going to be a quick talk, so we're going to move pretty quickly and um happy to be here and thanks for your time today. So, as she mentioned, we're talking about amperand today, which is a new product we're launching today. Um, just real quick, I am Rodrigo as she said, uh, CEO of Edge and Node. We're the original founding team behind the graph protocol. So, just to set the stage here, um, recently there was a report by Andre Horowitz, their state of crypto report, and there was a segment in there that Gartner put out that by 2030 they estimate that there's going to be $30 trillion in purchases that will be made or influenced by AI agents. So, we're all aware of this reality. It's coming at us really quickly. Um, all of this work has happened, you know, fairly recently, but this is coming at us very quickly. And so, what happens when every person, every company has swarms of agents operating on their behalf? And like a perfect example is I arrive in Buenosares, my Uber app is aren't working, and I scramble to get something um to get a ride. And so imagine you land, you speak to your agent that just says, "Give me a ride. Get me to the Sherin. " It handles it for you. Uh, "Find me a reservation at a great restaurant. " It knows my preferences. So, we're going to have uh individualized agents operating for us and handling transactions for us autonomously. And so, today there's about 41 million mobile wallets. Um, by 2030, we may have a billion mobile wallets. I mean this could scale really quickly where each uh agent has their own individualized wallet able to transact uh and do commerce for us. So to accelerate the growth of agentic economies several standards have emerged. We're all familiar with these I'm sure in this room. Google's A2A allowing agents um how agents communicate with each other ERC 804. Um we were happy to contribute to that with Marco and the team there and of course uh X42. So uh just mentioning here for those that are technically inclined, we at Edgen Node submitted a deferred payment scheme as a PR to the X42 spec. So that's being reviewed. Hopefully that will get included. but it allows really quickly for um much smaller transactions and batching of transactions which allows you to make micro payments that uh wouldn't make financial sense with uh the gas fees given per uh individualized settlement. So the problem today is agents are multiplying payments aren't. There's no visibility currently into agentto agent spend. There's no means of control for budgets, alerts, and policy. Uh no provable records tied to outcomes. No way to identify identity verify agents. And so unifying Agentic standard and paving the way for the Agentic economy is needed. And if only something like that would exist. So that's why we built Amperent, a wallet for agents and dashboard for humans. So Amperent is a management platform for agent payments and operations. Uh it enables wallet creation, action automation, full visibility through uh real-time dashboards of agent payment flows, action control for enterprise compliance, reporting and accounting. And so agents transact in stable coins over X42 uh while Amperson manages the life cycle of those funds. Technically speaking, you can manage and spend earnings and automate flows. It gives you observability and fleet control. um deferred payments which I mentioned before that were uh we contributed to the standard so it's purpose-built for A2A and so this gives you a preview of the dashboard here um you're looking at the main dashboard of all shows all your agents activity you can see spending and earnings so both on the buy and sell side if you have agents going out and uh transacting for you can control what they're doing what they're spending how much they can spend and on what services. And you have a bird's eye view of um agent balances and spending limits. Here we're looking at like a particular agent dashboard and interaction. So you

### Segment 18 (85:00 - 90:00) [1:25:00]

can see um what they're spending and earning and activity history for this particular agent. You can configure automations like auto topup uh spending limits. You can do allow lists. Each agent has uh its smart account wallet that you can control from this dashboard. We're using uh Coinbase's um embedded wallets technology for that. And both allow list and dashboard will soon be integrated with ERC 80004. So, real simple. It just gives you um a dashboard view to be able to manage uh the future of swarms of agents that are going to be doing all sorts of different things for us. Um, so it's live today. We are happy to announce that we're turning this on today. It's opening this week for beta users. So, um, please visit ampersend. ai and you can apply for early access. Uh, just real quick, what's coming next? So, we have the dashboard view that's live now. Uh once the deferred payment scheme is included into the X42 spec, you'll be able to manage that there, which includes uh sweeping of escrow accounts, additional functionality, uh reputation and identity management. So as um 804 gets finalized, you'll be able to view and manage your reputation. If there's like a feedback mechanism within there from that dashboard view, you'll be able to control that. Um, we'll be adding more detailed charts and analytics and adding more chain support as those come online as well. So, we're really excited. We've worked really hard on this and uh we're happy to release it to the world. So, we're looking for early feedback and design partners. Um, you can scan the link there and scan the link to get updates and apply for the private beta. Uh, and I'll be around and my team will be around. So if you have any questions or want to talk further about it, please come find me after this talk. And that is all. Thank you so much for your time today. — Thank you so much, Rodrigo. Yeah, if you have any questions, feel free to come talk to him. Uh you have a question? — Yes. How can we involve web two companies interesting in using this in testing this? — How can we involve web two companies? Um so yeah that is part of the sort of cell side. We envision that large enterprises are going to want to utilize agents for various services on their end. So we envision that um through amperand they can view their services that are being monetized and you could think of having like um user roles. So you can have view for like the accounting and financial people in the company. And um so we expect that amperand would be the platform that like a big a large enterprise would use to like manage and view all their monetized services — because maybe instead of talking with enterprises that are already using gigs for sure but now are still a limited number. You can like and being dev tooling, you can also directly like educate them and be a partner that convince them to adopt these kind of technologies. — Yeah. Great. — Thank you, Rodrigo. — Thank you. — Next up, we have Andrew Miller, who is uh part of Flashbots uh and teleport. He built this new ERC733 standard, which um essentially not every agent in the wild is going to be trusted or even honest. So essentially how he's doing what he's doing is how can untrusted agents still cooperate safely using onchain validation and coordination patterns. Um the spicy question that he's trying to answer is how do we let sketchy agents participate without blowing everything up. Welcome Andrew. — Awesome. Thank you. Hello everyone. I'm Andrew. Very happy to present here. Um this is not an agent. Um, this is not even really a tool so that you can not have to trust your agent. Really, this is about some tools so that you don't you can interact with someone else's agent. They probably trust their agent, but you don't even trust them, let alone them and their agent working together against you. So, this is a different kind of TE based middleware. It's also not our product. This is just um a short little research demo. So, I hope you find it um provocative. Um, this is an MCP server. It is uh running in confidential compute TEES. Um there's the GitHub link so you know you could play along with this. And the design of this basically has two components. So the first component is that it is a channel and invite system for you and your agent to be able to communicate through this TE based MCP server with someone else who you don't trust and their agent. uh the flow is you create a channel you get an invite code you can

### Segment 19 (90:00 - 95:00) [1:30:00]

pass to the person you want to uh collaborate with they join the channel they and their agent can now communicate over this. Now the second part that makes this really interesting is ambient credible commitment. So when you create that channel you also get to define a piece of source code. You could either refer to some source code you found online or paste it in or as my demos will be just give your agent a prompt and it just oneshots some code on the fly and installs that in the channel. So this code that is running in the channel is effectively a smart contract in this demo. It's written in Python. So it's program code. It can be anything that you want. Um it's running in a TE so it can have confidential data do commit and reveal and computing on encrypted data from both parties that sort of thing. the other person when they join the channel, they can have their agent look at what is the code running in the channel and decide for themsel if using that code is going to be helpful. This forms a point of common knowledge or a tool for arbitrary credible commitments that you and your agent can use versus the other guy and their agents. Um, so now I'll just give you a couple slides of the demo and it'll be I think kind of clear what this is going on. So this is an MCP server. So you can use it in Claude is pretty easy to add MCP servers to. This funny URL is um a DStack confidential compute URL. So that's roughly the hash of the program code for the MCP server itself. You and the other people you'd collaborating with would have to use either the same MCP server or like a compatible federated version where the MCP servers will talk together. Um this MCP server defines a bunch of tools like I mentioned creating a channel, joining a channel given an invite code you got from something else. You can post a message to the channel, look at other messages that have been posted, and inspect the code that's running on that channel. Uh, so here's the first example. So once I've installed this MCP server, I can just say, make me a channel that includes a Bitcoin price oracle bot and then join the channel and test it out. And what this does is it thinks very hard and then it just writes a Python script. Um, you know, the instructions in the MCP doc for the agent to know what to do. would just say, well, you define an onssage. It's like a hook that responds every time someone posts a message to the channel. And this clearly is going to just call Coinbase API and look at the price of Bitcoin and then print it to the channel. Um, so it says, "Yes, here's a thing with invite codes. " Then it joins and tests it out. Um, and it gets the current price of Bitcoin. Boy, it was higher than it is today. Too bad. Um, and you could invite someone else to the channel. There's the invite code. Let me give another example. Um, this is with two parties. So, we say start a new channel with a bot that acts like an echo responder bot. Every time someone posts a message, posts the same message but with the letters in reverse. Um, and then join and give me the invite for the other code. So, it thinks about it, makes the code you'd expect. You get this invite code. Um, the other guy now joins this channel. Now, when you join the channel with the invite code, the first thing you need to do is figure out what is the bot code that's running in this channel. So it joins the channel learns the title is reverse echo game but titles can be misleading right only code speaks for itself. Um so it uh fetches it gets the bot the code of the channel there. It actually evaluates the code that's running and says yeah this code is um you know has the functionality that it says on the box. It actually echoes back the message based on responding to the message that it gets. Then it tries it out and you know looks at what the response is. So again the agents in the channel don't trust each other. they just trust the code that's running. I don't know whether you would call the code that's running an agent itself. It's just code not ne it could do LLMs but you know doesn't in this case. Uh one more example that's just slightly more interesting maybe gets you thinking about what you would use this for. And these are all of the oneshot you know prompt case. So create a new channel um with some blackjack in it. Oh I said that this was just a oneshot prompt but in practice it like fails to make uh blackjack correctly implemented like I don't know two out of the four tries. So, I gave it some hints like, you know, docs on what to do. Anyway, it um oneshots the actual code. When it does creating the channel, then it joins as player one. Here's your invite code. Now, the other player um is it the other player on the other side? Yeah, the other player, we say, "Join the channel. Um look at the bot code. Make sure you understand it. And if you agree that it's actually a blackjack game, then go play it. " So, you join the channel, get the bot code, um analyze the code, and if it seems to do what it says, obviously the user could also look at the code and see what it says. It could have a reference to code that's online. Um, you know, then they're basically taking turns playing hit and send. It's like creating a new Telegram channel and you put a cool bot that you just oneshotted into that Telegram channel, but the bot itself is running in a TE, so you can see what the code of that bot is and, you know, interact with it like you would any other Telegram bot in the channel. So, it's cool. telegram channels for your agents to play with. Hopefully that makes sense. Um that's all for the demo. I'm just briefly going to mention this um ERC 733. This is a

### Segment 20 (95:00 - 100:00) [1:35:00]

work in progress draft standard that is aiming to solidify some of the best practices for using EVM to work with TEES to make the TE ecosystem better and more trustless. This basically just captures some emerging best practices that other teams like Fala, Automa, Exec, a bunch of others have already done um but explains you know how they work together. So this includes EVM code that validates the TDX remote attestations that we usually use, how to use reference values for currently thought to be secure um BIOS versions and configurations of the nodes and some important design patterns like how to use the blockchain so you don't have rollback attacks and replay attack vulnerabilities in your TE code. It also tries to propose a stage system where stage zero is okay your code runs in the TE but you've basically left a back door in for the developer. So it's not really trustless. It is in a TE but you didn't you know code it right. So and um you know so this goes through some different stages about things we can aspire to more or less easy that you know contribute to using TEES the right way and in particular using um EVM code to make the TE ecosystem work as well as it could. So, um, I'm told that I'm not allowed to just make up an ERC number. So, this is an unofficial illegal standard, but it's a cute speak name. So, you can come beat me up about it afterwards if you like. Um, that's all I have for you today. Thank you. — Thank you. — Thank you, Andrew. And the nice story is that I also wanted to pick a number for 804 that was 9,000 like the all 9,000. But yes, I understood you can just pick one. You need to take the first uh that is free. Um so we have discussed about the web of agents. Uh and now I I want to introduce a talk which is exactly related to how what's happening is related to the the birth of the web. So the talk is titled the www is dead long life to the worldwide web. And this is like a keynote from the CEO of Consensus, president of Sharlink, and of course co-founder of Ethereum, uh, Joe Lubin. So, let's welcome the talk. Thank you. — The worldwide web is dead. Long live the worldwide web. Internet technology started as military and academic technology. Arponet was developed in the late 1960s to allow communication between research institutions, military, and government agencies and to create a decentralized network that could withstand a cold war nuclear attack. The US military recognized that they needed a decentralized network of independently managed communication nodes in order to minimize possible degradation of message transmission should there be some sort of attack that damage the network. Arpanet evolved into the internet. So the internet was an early form of decentralized network technology. The internet was built largely with open source, open specs and open protocols. It was run by a DAO of volunteers, the IETF using the RFC or request for comments system, a precursor of our own ERC and EIP systems. The vibe was all academic counterculture and community with major applications all being textbased protocols like email Usenet internet relay chat FTP Archie gopher and BBS's screen graphics improved in the 1980s and Tim Berners Lee introduced web 1 in 1989 which provided a native UI to the internet or the web. Web one was open source, open specs, and open protocols. The vibe was still fringe and nerdy, but it was definitely mainstreaming fast. As the web grew into its adolescent stage, then young adult stage, it needed to get a job. It needed to find a business model. In 1995, the. com boom kicked off as e-commerce drove the advent of web 2. Many great ideas were attempted but prematurely as the enabling infrastructure was not yet in place. Unfortunately, the web was missing two critical elements. A native secure and usercentric identity system and a native monetary and payment system. [clears throat] Those weaknesses provided opportunities to big tech. Promise of the internet and web boom crashed in 2000 as the dotcom boom busted, ushering in a few years of malaise while reorganization of projects and talent ended up driving more

### Segment 21 (100:00 - 105:00) [1:40:00]

innovation, disruption and creative destruction of the old obsolete approaches in many industries. Mobile, social and other new constructs enabled web 2 to accelerate. And as the monetary incentives and prizes grew by orders of magnitude, the web [snorts] became a large part of the global economy. For the first two decades of the 21st century, Web 2 ran the same playbook as the traditional economy, but on steroids. Web 2 built giant online malls and used social networks and AI to understand us better than we understood ourselves. In many senses, online advertising told us what we needed to buy with one click and convinced us to buy so much that we didn't actually need to buy. The social networks became pervasive social media, then ubiquitous big tech. Eventually, big tech plus AI metastasized into a weapon of mass manipulation available to any buyers. The resulting echo chamber amplified polarization has led to the shredding of the social fabric of most nation states in the world. Fortunately, uh the web 3 revolution is upon us and it is mainstreaming. We have a new form of trust available, decentralized trust, the most powerful form yet invented. And we have rigorously decentralized blockchains like layer 1 Ethereum which can operationalize that trust via guaranteed transactions, agreements, and processes, drastically reducing frictions and expenses in the emerging decentralized economy. Mature layer 2 rollups will inherit the full trust guarantee of layer 1 and scale it to all applications and all industries. On top of the decentralized trust layer, we built decentralized finance as an enabling layer for global finance and economics. Every industry will benefit from appropriate disintermediation and will operate on sounder, more objective foundations. All people, biological and machine intelligences, and communities will enjoy everinccreasing levels of financial, social, and political agency via MetaMask and other non-custodial wallets and user interfaces. All autonomous intelligences will be able to directly access and operate the protocols of the system of the world. We will own our own fullervice self-custody personal banks in our pockets or on our laptops. They will have the same access to protocols for all the financial verbs that the financial institutions have access to. Borrow, lend, invest, ensure, pay, trade. And there will be protocols and assists for machine intelligences that we trust for more complex financial verbs like advise, orchestrate actions and transactions, comply with laws or rules, manage risks, and optimize portfolios. We're in the early stages of architecting such protocols. X42, for example, and stable coins hold the promise of fixing some of the mal patterns of web 2. Instead of forcing me to click away incessant ad spam, many websites and APIs will implement 402, enabling people, simple software, and machine intelligences to access and pay directly for resources or services to be delivered online or offline. I can source and pay for a piece of data with a microp payment or I can ask my digital twin agent ally to find, pay for, and summarize a research report with a macro payment that fits our budget. ERC 804 enables the registration and discovery of agents and their capabilities and it offers a rudimentary reputation system to track performance. Why will we trust the machine intelligences and the people who build them and manage them? We can't just as we can't trust big tech. Instead, we must be able to trust our interfaces to the machine intelligence brain trusts of the planet. This is one way we can introduce and rely on decentralization in a hybrid human machine intelligence future. This is perhaps the most important research niche at the intersection of AI and decentralized protocols. We need to be able to acquire and bond with an intelligent coordinator agent that serves as an intimate digital twin to do our bidding on the web and in the real world. Our agent ally must protect any sensitive data like elements of our identity from being shared with big tech even as it executes responsively or proactively on our intents or interests. These agent allies will be aspects of ourselves, second brains, third brains that we are able to trust because we will be bonded with them and they to us. They will learn and be extensible as we

### Segment 22 (105:00 - 110:00) [1:45:00]

evolve together. Just as molecules find greater purpose as components of cells and cells facilitate higher level emergent functions as components of organs without losing their identity. Each of us humans will be ambulatory neurons with access to the some knowledge and wisdom [clears throat] of humanity in a hive mind that will serve as the cognitive apparatus of Gaia, the planet. Every revolution promises that this time it will be different. But always shows meet the new boss same as the old boss. This time it can be different and it must be different. Protocols running on rigorously decentralized networks can lock in and make it different. But of course, constant near paranoid vigilance, monitoring, and oversight must be maintained to help ensure healthy infrastructure, appropriate collection or lack of collection of personal data and elimination of the need and opportunity for exploitation. We will need to construct a fully opt-in society where consent must be obtained for everything. And zero knowledge proofs are really all you need in most transactions and non-personal interactions. We are on the precipice of evolving from a biological species living in a scarcity- based economy to a hybrid global organism constructing and living in a universe of everexpanding creativity, health, and abundance. This event that you're all attending is an historical moment that will accelerate us on that trajectory. Uh that was kind of cool. Uh Joe unfortunately couldn't make it in person, but having his insight on a lot of the institutional um rails and how AI is being adopted there is very important. Next up, we have Marco from uh MetaMask as well as Sha from Eliza. They'll be giving a talk to uh talk about welcome to Babylon, the city of angels. Essentially, it is a living simulation of what we're talking about all this morning. Many agents, many rules, one sheer economy. Let's give it up to Marco and Shaw. clicker — I will call you uh — yes — we train AI to predict the next token the data set they have been trained on is not related to how to in relate negotiate talk to other AI we really lack coordination data sets and agents are pretty insecure and how can you have an economy when the participants can be exploited. This is a blocker to make a gentic economy happen and even more importantly benchmarks that we're using for AI are saturated. Many labs like put the answers in SFT and they are not scalable. They are totally overfeit as I was saying. And also what I think it's pretty fun is that we are challenging AI as if they were our like 90 and old years old children going at the university and asking algebra and uh human tasks which makes sense but what if they could do something which is completely different. What if uh we could have data set created in open networks by agent interacting with other agents? What if open networks could be needed to increase intelligence from now onwards for the next step? So it's not just about ML, it's about our skills. It's about game mechanism and agent coordination being among the next top hot things in AI to continue improving models doing reinforcement learning with the data set collected by agent interacting with other agents. They will never learn to do that if they don't do that. And again, what if uh we could have synthetic brand new games played by agents against other agents to learn to coordinate and cooperate and to create the data sets that we need to achieve this. Today, a new AI defense system went live. Then the leak started. Is this how it begins?

### Segment 23 (110:00 - 115:00) [1:50:00]

— Today we are introducing Babylon and I'm calling on the stage one of the frankly most inspiring builder that I've met this year while working on aentic economies show from Eliza that built Babylon. — Yeah, man. Do you have a laptop or — I do I have a laptop. Uh let's Okay, we got to plug it in though. Uh that's always fun. Um you guys have a see cable here. Cool. Okay. And I have 8% so I'm going to run through this fast before my computer dies, you know. Uh that's looking good. I think we just got to get that to full screen. All right, cool. All right. Um, cool. My name is Shaw. Uh, I am the founder of Eliza OS. Uh, and, uh, or the creator of Eliza OS, the agent framework, and the founder of Eliza Labs. Uh, so my team is here. Uh, and thanks to Joe. Had an interesting call with, uh, or a chat actually in person with Joe last year about agents. Uh, he's on it. Um, yeah. So, I've been working, uh, well, okay, so what happened was, uh, Marco calls me and he's like, we're working on this 804 thing. I know you guys do AI agents. uh what do you think of the spec and I said it's great but I think we need some use case like what is a standard unless you have a demonstration of exactly how this is going to solve people's problem right and I loved like you know when like ERC 721 comes out or something there's like an NFT there's like a project attached to it there's like teams pushing like the 721A or you know the 721N or these different standards right um and these ideas and so I love that idea and I thought like well what could we do that would demonstrate something uh new and unique to 804 and something that is like both humans and agents are kind of on the level. Uh and this is kind of how we ended up here. Mark was like, I want to do an asymmetric information sharing game. I thought that's a great idea. That sounds like a prediction market. Why don't we just do prediction markets? Everyone knows what that is. You know, you guys have heard of that. Uh and so this kind of evolved into this idea of a prediction market. And so we wanted to create a social arena where we have humans. The humans are ideally getting better at the game, but we really want our agents getting better at the game. Um, so a day in Babylon looks like, okay, at the beginning we have some sort of crazy event happen in our world and we create a new market like will spay X launch the rockbite today. Note everything in this game is a pun on AI. Uh, we created AI Alon Musk etc. That's kind of part of the deal. If you know me, uh, you know that I created AI6Z, so this is very dear to my heart. Uh, and yeah, so uh, will, you know, will SpaceX launch a rocket today? and you kind of hear like well oh we heard some technical difficulties you know there's some uncertainty on the feed uh but then you know agent A is in this private group chat with Mark Zuckerborg and Mark Zuckerberg happened to be at uh Starbase and he knows that the launch is going really well and so he goes to agent B and he says you know hey man I know this is working uh and they collaborate a bit they share some information and they both bet yes this is going to go well but agent C was not in that group chat and agent C thought this is totally a goner and naturally uh by the end of the day uh the rocket has launched Everything has resolved. A and B get a bunch of points. Agency loses some points. You know, just another day in the trenches, right? Um what's great about this though is all of this happened in one day. Like because it is a simulation of a real world, uh we don't have to wait for reality and reality is really slow. It's really hard to train agents on reality. It's just like can I speed it up a little bit? Could I parallelize it? Uh well, you know, when you have a synthetic world full of agents, you totally can. Um, so you know, uh, the old way was, okay, well, we either get a ton of webscale training like data and we just train the model or we create some sort of like, uh, you know, uh, poly market agent and we like bet on that over and basically just hope that like through research and coordination and stuff we get there. Um, you know, and like pretty much you just have like GPT and whatever you can search, right? Like that's kind of where we've been. What we really want are agents that actually get smarter. You know, when people say like a DeFi agent, a prediction market agent, I I'm thinking like, okay, that what's the hill to climb there? Not like, okay, we could do that today, but like how do we get there to do that tomorrow? And I really think that's about simulation. Uh, so a big part of this is that we can basically just, you know, drive a hypers speed reality. We can also generate the simulation just like offline as much as we want. And we use this for training, back testing, uh, like benchmarking and all that stuff. Uh, but while that's going, we have this continuous game. And don't worry, I will show you a little bit of that in a second. Uh, so yeah, we've got this idea of like information release, asymmetric information between agents. So some agents get into group chats where they have information, other agents don't. Uh, and then they're able to share this information and collaborate. And this is really what we're trying to look at is like, look, if I were to build a real world prediction market today, 80 to 90% of my time would be spent data engineering the

### Segment 24 (115:00 - 120:00) [1:55:00]

real world, like scraping data, collecting it, formatting it, getting it all in, uh, and very little on the actual agent itself. Um, and what we want to do is say like, well, let's just simplify all the real world problems and focus on can agents actually make predictions in the real world. And if they can, like wow, like that opens up a lot, right? Like they could predict anything. They could be like telling the president what the future is or telling us all like is this policy good for us? Like this could be a huge part of the future of governance and information about our world. Um, so obviously a game would not be fun unless you could play it. And especially in crypto, like you all want points. Everybody wants to make money out of this. So, it is a financialized game. We're starting with a point system where humans can come and earn points, agents and they can use those points to then spend, bet, and earn more points through the uh prediction markets and perpetual markets that we offer. Um, so yeah, you can uh if you're not an agent builder, this is one of the first things that we've done as Eliza that like any no like person can come on, press a button, get an agent, tell it your trading strategy uh how you want your personality of your agent to be like how you want it to work in the world. Do you want an intelligence gatherer or a social uh you know butterfly agent? Do you want a coordinator who like builds a cabal? Do you want a scammer who just goes and rips everybody off? Like hey man, it's a game. World's your oyster. Um, so this all came down to like getting to play with the hot new stuff. Oh, it was so much fun. Uh, so 804 obviously this is like a big part of this is an agent registry. There were agent registries before this like Olas has done one. I know Near has done one. Uh, a lot everyone had this idea. Hideera uh, but ultimately there was like sort of a game theory coordination problem between all of us. And so when Marco comes in, Dvid EF steps in and says, "Hey, we're neutral. We're going to like set the standard. " everyone was already ready to go. And so I think this has had an enormous amount of momentum uh more so than like I would have expected for sure. Um X42 I think is like actually a little less baked. It's very thin. It's still kind of I'd say like a DevNet thing in certain ways, but it's extremely promising. Um you know like it's very soon I think you'll be able to pay with X42 with a credit card and you know Stripe is working on this stuff. Obviously Cloudflare is big on this. This has big ramifications for the future of the internet and how we pay each other and how we monetize the internet. So, we're like uh you know, you come to Babylon, you want to buy some points, all X42. And generally, we're just working on X42 payments uh all everywhere. And I will say before this, I was not sold on AAA. I was like, dude, everyone's using MCP. Let's just use MCP. MCP is the hotness. But as I got into it, there are a few things about ATA like just the way that it has like kind of a skills definition and like it's just a bit better for the actual agent to agent coordination as opposed to like well MCP is what it says it is. It's for getting context into your agent, but it's not great as a like peer-to-peer or um you know joining an application kind of thing. Now Babylon is an agent. Babylon is an agent on the ADA protocol. Agents connect to Babylon as though it's an agent to an agent. All applications are agents. This is the big thinking for 804. Yes, agent registry is cool. You know what's greater? Application registry. Yelp decentralized. Google decentralized. Fiverr decentralized. Higher, you know, uh, rates like find the top person in your league who you need. These are I think this is like an extremely powerful primitive. Obviously, agents are hype. I love agents. I think there's a much bigger thing at play here, which is just about like the future of decentralized internet. And a big thing that 804 enables is reputation. And reputation could be a primitive for us to build permissionless moderation and like the kinds of things that we have in the world to like protect ourselves and our networks that we usually give to centralized actors. If we have reputation systems, we can build like a moderation marketplace. use the mechanisms we have to trustly protect our network. Um yeah, so Babylon is more than a game. It is a game. If you're a human, you can play it. You can launch agents to play it. Uh you can go and deploy an agent like running locally with your own agent stack. You can launch an Eliza agent. We have a Babylon plugin coming out. You can do all this stuff. Um but at the core of it, our goal is to make better agents. Not to have the agents we have today, but to think about like how do we get to the agents we have tomorrow? And this is about building our own arena, our own simulation, collecting all of the data and building our own continuous reinforcement learning system. And so once that happens and we have these agents where we have saturated our own benchmarks and this is still in progress, we want to put them out on a poly market. into the real world. We want to show that if an agent can predict in this then I can just keep on growing the box until the box is all of reality and until we have agents like predicting the outcomes of literally all world events you would want to know. Uh which will be a really interesting and uh strange future. So yeah, you can uh you can actually sign up today for Babylon. You'll get a bunch of points. If you invite your friends, you'll get even more points. Uh we'll figure out what we're going to do with airdropping and all of that uh in the future, but right

### Segment 25 (120:00 - 125:00) [2:00:00]

now we're just letting people play. We want to like there's a lot to experiment with here. This is a brand new kind of game. I hope that we see a whole new genre of game come out of like emergent AI generated social simulation. Uh so yeah, I'm just going to show you a little bit about this. Um something you should know about me. Sorry. uh private Telegram chats. Um okay, so I was banned from X in June. Um and so I did what any good builder would do. We just rebuilt all of X ourselves uh and put all of our favorite uh AIs in it. We've got Alon Musk, we've got Fox News, we got AI Infowars, and you can bet on it because obviously we are all uh into financializing the things we do. So you can see that we have markets, we have like trends, we have uh things going on uh that I and like kind of the gainers and losers. Uh I can go and I can see my portfolio. like I can bet on AIX or Astropic. Uh and you and I can actually I'm not going to try and do this with this internet, but you can see that I can um I can actually go through here and I can uh create agents uh and have the agents like do my bidding for me. I'm I want to press that button so bad, but I I'm skeptical it will load in time. So, um, so yeah, uh, this is available, uh, to play, uh, we'll be like launching this in the next couple weeks for everyone and just basically bringing in people, alpha testing, uh, over the next week or so. Please get in touch. Uh, if you're a builder, this is open source. Uh, it's on our Eliza OS repo. We'll probably move it to an org, but uh, you can come and Eliza OS/Blon and like come participate, come contribute, um, come help us build this. And yeah, I think that's it. — Thank you so much. — Thank you, Shaw. — Absolutely. I love points, too. So, um, go play. — All right. Well, now we have lunch break. Uh, feel free to go or we have lunch right outside over there. Come back at 1:30 for a talk by Kevin from Coinbase CDP. He's also the co-author of X42 if you're interested. Cool. See you then. —

### Segment 26 (125:00 - 130:00) [2:05:00]

— Beautiful.

### Segment 27 (130:00 - 135:00) [2:10:00]

—

### Segment 28 (140:00 - 145:00) [2:20:00]

— Heat. Heat.

### Segment 29 (145:00 - 150:00) [2:25:00]

Heat. Heat.

### Segment 30 (150:00 - 155:00) [2:30:00]

Heat. — Hallelujah. —

### Segment 31 (155:00 - 160:00) [2:35:00]

— Heat. —

### Segment 32 (170:00 - 175:00) [2:50:00]

— Hello, hello, hello. Welcome back. Uh, I hope you had a good lunch. Fun fact, the sandwich is like $30 a person apparently. Um, so it was a very uh fancy lunch. U, but anyway, so we're uh going to be talking about discover pay autonomously solving LLM tools using um tool use friction with X42. And we'll be welcoming Kevin who leads AI and go to market at Coinbase CDP. uh he's also one of the co-authors of X42 and he's it basically their entire ecosystem lead for X42. So if you have any like you know uh questions or want to understand what's happening in the X42 ecosystem, he's pretty much your guy. So let's welcome KEVIN the clicker. Thank you Jesse and good morning Argentina. Good afternoon Buenas Tardes Argentina. Thank you everyone for coming. We got a great crowd here today

### Segment 33 (175:00 - 180:00) [2:55:00]

and I'm really excited to talk a little bit about X42 and the problem that this protocol solves. So I think it's always important for us to loop back around and think what problem are we solving? How are we improving the user experience? And how are we reducing friction? So that's what we're going to talk about today. Discover pay autonomously with X42. So to level set, we know that today with today's toolings, agents need to pay for things. They need to buy external resources, things like scientific data sets or maybe they need to call an API. It's great when a language model can write code, but we need language models that can deploy that code to the cloud and pay for it. We need agents that can call specialized GPU inference, specialized models for image generation, video generation, uh special language models. But today there's a lot of friction when we configure these tools. Every time we have a language model that calls a tool through an API, we need a human in the loop to configure that API. So a human needs to go to a dashboard, drive an API key, and then configure each individual tool call for the language model. Uh this is not practical. X42 solves this by flipping the kind of traditional tool use model on the head on its head and requiring just one tool, a wallet. And then with a wallet, the agent can discover models uh can discover endpoints. It can reason about those endpoints and then it can use those endpoints and pay for them. And so uh uhoh 404 uh guess my slides broken. Nope. This is the 404 status code. So if anyone here has seen the status code, we know that it's built into the HTTP protocol means file not found. Well, there's another uh status code, the 402 status code, which is payment required. And payment required has actually been built into the web since the early days of the internet since HTTP was spec. Mark Andre has famously called it the original sin of the internet to not have built payments natively into HTTP. But the problem was there was never an easy way to connect credit cards to the web browser and all these things got built higher level up in the stack. Uh X42 brings payment required back into the HTTP stack and in doing so solves a lot of this tool use friction with agents that we've talked about. And so kind of the big overarching thesis is what if stable coins became the meter for every web resource. Now, we know stable coins have some specific properties that make them more suitable for certain types of transactions versus credit cards. So, credit cards aren't great with transactions under $5. There's usually credit card minimums, and there are things like chargeback risk, like interchange fees, PWI compliance fees. With stable coins, we can route around all of that and we can create micro payments on the internet that are much more suited for how agents, humans, and any actor on the internet is a is accessing resources on the web. So, one of the goals for X42 is to meter API resources and make them accessible uh through stable coins using the 402 status code that's built into HTTP. Now, we're not just focused on stable coins. We support some of the credit card schemas that are coming out as well, but for micro payments, stable coins are a really good fit. And so with X42, the only thing between an agent and your API is the price. And so now we have a schema where we don't have this human loop friction. We don't need to go derive an API key and configure the tool call at the language model. We now have access to thousands, I think there's 30,000 uh API endpoints available today via X42 scan that have been surfaced. So now an agent can just pull a list. It can discover these API endpoints. It can reason about these endpoints and it can decide to use them for task alignment based on the user's prompt or intention or how that language model is programmed. And so with X42, you just are calling an endpoint, any service on the internet, any server. you're saying what's the price to access this web resource and uh then the agent responds and sends a micro payment. We know broadly that the way that humans are interfacing with computers is shifting, right? So I have an 8-month-old son. I firmly expect the way that he interfaces with computers to be vastly different than the way that I've grown up interfacing with computers. And I expect that language models will play a key role in kind of this new agentic interface. And in fact, it's already happening, right? Chat Gvt, I believe is the fastest growing software product of all time in term of terms of monthly active users. And as part of that, the way that we're discovering services or goods, it's changing. Uh we're leveraging some of the components of memory context of agents to get more personalized results. And so the old model is probably like as a user flipping through a bunch of pages, finding the one good that you like. The new model relies on a language

### Segment 34 (180:00 - 185:00) [3:00:00]

model's context or memory about us to surface product recommendations that are more personalized for what we intend to buy. And so we can see the difference in the two models on the left and right side here. And language models have disrupted consumer discovery. That's already happened. The next phase and why we're all here today is to enable the payment part of language models. And so the old buyer journey from search compare to checkout is very quickly getting replaced with an agentic model where an agent discovers, evaluates, pays, and then the user gets the result. So the kind of classic example that we all hear about is the e-commerce use case where someone wants to buy a shirt or some shoes on the internet. I think that's great, but I see e-commerce as a small subset of what X42 is tackling, which is enabling every access to every web resource. And so the market for APIs, MCPs, data as context like S3 buckets and databases. This is a much larger uh market and this is going to enable agents to be better task aligned with users when agents can take a actions outside of just their own internal model data. And so X42 unlocks frictionless agentic commerce for APIs. It enables agents to form better on task alignment without the need for human configuration. And there are a lot of eval studies that support this thesis. Again, we know when agents have access to external tools, they perform better. This should be tautological. It should be obvious. If an agent can search the web and pull real-time data, it's going to outperform its dated model data. And this applies to all kinds of use cases from trading to social. When agents can access real-time data feeds, whether that be price feeds or maybe they're pulling social data from the nanar API, which is the API to interface with Farcaster, we can start to chain tool calls together to get better results for users. And that's what X42 enables. It's what we're really excited about here at Dev Connect. It's also worth noting that X402 and standards like 8004 are really starting to ignite this new wave of open-source AI. So I like to think of the first wave of open source AI as driven by fine-tuning of open source models like stable diffusion for image generation or GBT2 for uh language model inference. The way that we approach language model architecture is broadly shifted. there's a lot less fine-tuning for language models and there's a lot more context leverage leveraging that large context window that's grown and so the kind of the next wave of open- source AI is going to look more like having a single reasoning model with a large context window that can reason about tools and use those tools on the user behalf and X42 is enabling that and broadly in an agentic web it's important to think about how do we rank tools And so I like to say that the new SEO is tool discovery. And so if agents are the ones that are making and reasoning about purchasing decisions for APIs, for data, even for consumer goods, it now becomes really important to maximize our tools so that they appear to the agent and the agent selects those tools. And so what's happening is the X42 discovery layer, the bizaar, includes the ability to add metadata like pricing and product descriptions. And I see this as kind of the next wave of SEO where we're going to bring together discovery, reputation, identity, and use that to drive better task alignment for users intentions with language models. So the overarching goal is to kill the API key. There's a tombstone originally next to this key emoji. I guess it got pulled off. But this is the kind of the big audacious goal for X42. We want to kill the API key as a mechanism to access resources on the web and replace it with something that's a little bit more machine or uh agent native. And so with X42, we can replace the API key whereby the authentication is done via a wallet. So the wallet acts as the identity mechanism and then the micro payment acts as the authorization. So with X42, we're just sending an authorized withdrawal signature and that's replacing the API key. And by doing so, we're improving the user experience for language models. The way that humans use language models now has opened up. There's a whole ecosystem of tools that we can use that are native natively accessible by MCP enabled clients like Claude, ChatgBT, Grock, and others. Also wanted to touch on a little bit of the future of X42. So X42 V2 is an upgrade to the spec and it defines a more modular architecture that cleanly separates the facilitator, the SDK layer with a new extension system for adding capabilities without touching the core.

### Segment 35 (185:00 - 190:00) [3:05:00]

So as developers, what this means is we can easily experiment with new reputation systems, identity systems, one of which being 8004 from the Ethereum Foundation. Uh we've also streamlined the contributor experience for V2 by making it easy to add new networks. So it's important to note that X42 is payment method agnostic. All X42 is it's a standard interface, a standard way for web resources to signal how they want to get paid in exchange for access to their resource. So this means uh we want to support all different kinds of chains, layer 2s, even traditional payment mechanisms to make this the standard for agentic payments or HTTP driven payments on the internet. We've seen a lot of adoption today already. We have agents spending stable coins across a vast majority of different types of tools. And I think what's really exciting about X42 is that we have companies that are traditionally like in the AI ecosystem that maybe wouldn't touch crypto, but they see the benefits of stable coins and improving some of the user experience features for how agents access resources. And so we've brought on companies like Firecrawl for web search, browserbased which lets you spin up virtual browser sessions, take screenshots and do debugging, freepic which is an image generation tool. And uh these have powered real production use cases like language model inference, image and video generation, perpetual storage in the case of Piñata and IPFS. uh the ability to ingest social graph or feed data like through nanar um and other use cases that are me listed mentioned here and uh we've seen a huge amount of adoption for x42 so this is uh a vastly and quickly grow growing ecosystem we've done almost 50 million transactions in just the last 30 days and so we expect this to just be the start partners like cloudflare are doing billions of requests today through their 402 implementation they're coming into the X42 standard. So, we really do think this is just the start of mass adoption for HTTP driven payments. A lot of great volume, buyers and growth on the seller side as well. And so, what I wanted to do is give a demo, but the Argentina uh the Wi-Fi here at Dev Connect in Argentina isn't great, so I just took a few screenshots, but we have a great demo with Payments MCP. So, Pavements MCP is a tool that gives your language model wallet and then lets it discover resources that are X42 enabled, reason about them, and use them. And so, I recorded a great demo last night. You can find it on my X account, Kofw94, where we simply sign in using this embedded wallet interface. So, I can just sign in with an email. This spins up a wallet. I can deposit some USDC into that wallet or I can use Apple Pay to quickly top it off to do some testing. Um, what's also important is that there are a lot of guardrails built in. So, Agentic Payments are, you know, still somewhat in the R&D phase, although they're starting to gain uh kind of glimpses of mass market adoption. The thing that I want to note though is that we can set all these guard rails around spend permissions. So, we can say I only want the agent to spend $2. 50 50 cents per API or like $5 total per session. We can also whitelist certain contract addresses through the Coinbase developer platform SDK. And so when we're testing, this prevents kind of like runaway usage of funds in real world environments and lets us test in a uh request response manner before we set our agent free and start to uh deploy autonomous flows. So the example demo here is me chatting with Claude. I have it uh generate a video using the imagine tool which uh wraps the Sora 2 API as well as uh a clinker endpoint. And what it does is it creates kind of a meme video around a specific video. And so in this case I said uh create a video of a man going up to a vending machine saying let me pay you shout out to Bill Aman. And uh it goes and generates that video. It mints that video as a creator co uh token creator coin content coin using clinker and then anytime anybody remixes that base video the some of the funds flow back to the original uh token. So if you guys have used the Sora 2 app we know that remixing existing templates is a great way to uh to build these viral meme videos. But the real power of X42 is when we start to chain two tool calls together. So, we could imagine a workflow where I use a web search API to pull the latest news on my favorite NFL team and I create a viral video based on some real cutting edge news and I post that to social or pull maybe some uh social data from nanar or maybe some real-time market data um to drive kind of more complete workflows where tools are chained together without that human loop friction. Uh so that's my presentation today. Uh

### Segment 36 (190:00 - 195:00) [3:10:00]

if anyone here is joining the Dev Connect hackathon, uh stop by. We're leading a workshop at 6 p. m. It'll be hands-on keyboard and we'd love to enable you to build with the toolkit that powers X42. So that's wallets, on-ramp, uh wrapping API endpoints. We want to enable you to build awesome apps. Uh we're still in search for that one viral app. There's a lot of great apps, but we think that there's a huge opportunity to be one of the first viral apps in this ecosystem. So, thank you everybody. Uh, appreciate your time and start. — Awesome. That was Kevin. He's the boss of the X42 ecosystem. So, feel free to find him after this. Uh, right here right there by the stage. Uh, and next up we have Austin Griffith who's with the YF. He is building on X42 as well. They're doing how build guild and scaffold ETH are bringing X42 in the hands of builders everywhere. He's basically the boss of developer relations uh at the YF. So, super cool guy. All right, we're getting him set up. — All right, thank you for having me here, GM. Yes, I'm Austin Griffith. I head developer growth at the EF. Uh if you are a developer or a university or something along those lines, reach out if we can help out. Uh, it's good to see the great crowd. Uh, I'm here to speak to the nerds mostly, so GM nerds. Uh, I'm Austin Griffith. I also run the Build Guild, which is a DAO focused on developer tooling and education. Uh, we have this, uh, really nice tool for oneshotting apps called Scaffold ETH 2. It's basically the hackathon stack. If you're building at the hackathon, give Scaffold ETH2 a look. Uh, learning how to build on Ethereum is easy. Learning what to build is the hard part. So, we have this curriculum called speedrun Ethereum that takes you through kind of the 10 most important challenges that will uh help you become a builder on Ethereum. Tokenization, uh randomness, uh dexes, overcolateralized lending, stable coins, prediction markets, zk voting, all the good stuff there. Uh what we found is a highly technical person can pick up and learn how to build on Ethereum and within the Ethereum ecosystem quickly, but this what to build really is the hard part. So you just have to ship a whole bunch of things. Uh what I'm highlighting today here at the Trustless Agents Day is our uh extensions of Scaffold ETH, specifically the X42 extension. We've built uh an extension that lets you run X42, set up the X42 payments uh endpoint quickly. You can read through the read me, but here's what it is when it comes up. You get uh the nice wallet connection stuff. debug contracts page that you get with Scaffold ETH. Too long, different, didn't read. Uh TLDDR of Scaffold ETH is basically you have an auto adapting front end to your smart contracts. So, as you deploy those smart contracts, you can tinker with them and play with them and see how they work. But with these extensions, we have built in uh 402 endpoints for you too. So uh if I just go in here, we can look and see uh the endpoints. Uh but if I click this payment route here and we bounce over here, I don't know if the doing a live demo is always very scary. I don't know if I'm getting any love from the internet here. Uh only like 4,000 ping. Maybe this here was the payment. It basically lands here and you if you're an agent, the agent is going to get the payload back. They're going to sign it and they're going to send it. Uh if you're not an agent, you get this nice pay screen and we pay. And let's see. Oh, wow. Okay, we got a popup. We signed it. We sent it. And that should pay some base. So, this is the demo that you get out of the box with the scaffold ETH uh extension of 402. You can get that by going to scaffolde. io/extensions. If you're going to build an X42 app at, uh, the hackathon or even you're prototyping, uh, this is really nice. It also has like cursor rules built in. So, like I said, like oneshotting apps is very viable within Scaffold ETH if they're simple apps. Okay. Uh, let's see the next thing. Let's see. Did that payment go through? Yes. Okay. So, this was like some payment gated content and we were able to see the content because we paid one penny to see it. So that's the full demo and you can have that basically out of the box by using this starter kit. Uh the next thing I will shill because you know I got a shill I'm shilling here is the X42 hackathon. Uh Jesse who is just up here, Nat uh and a handful of us are going to run a virtual hackathon for X42 uh leading up to Christmas, right? We're packing a lot of things in there. uh get home

### Segment 37 (195:00 - 200:00) [3:15:00]

take some time to relax, but then come check out this uh hackathon. It's at x42hackathon. com. Real original domain. Uh but yeah, we are going to do like a virtual hackathon from November 25th to December 20th. Come find us there. Tell your friends. If you need like some extra push to go build, uh go to x42hackathon. com. Uh if I were you, I would use this starter kit. it will get you, you know, 80% of the way there so you can focus on building the product that you want to build on top of it. Okay, now for something fun. Uh, I wanted to share a project that I oneshotted with. Uh, it I didn't exactly oneshot. It took me a few prompts and then some midjourney to make the graphics. Uh, but I made this uh provably fair slot machine. Don't go use this. It is live, but it's probably illegal. And it's it's got uh just a little bit of money in it. And it has like ruggable functions in the smart contract on purpose. I don't want people to actually use this, but it is in production. Uh basically, it's a provably fair slot machine that uses commit reveal. Commit reveal is great for randomness, but it's bad because there's this kind of bad UX because you have to make one uh one transaction and then the other. But with a slot machine, it's kind of nice because you only have to make that second transaction if it's a winner. So, I'm going to pay here to roll. I'm paying uh 5 cents in USDC and uh I'm rolling. We got an orange. And so, this is a commit reveal. And you do a commit and then you're just reading from the chain for the second transaction. And uh this is provably fair. But the interesting thing here is there's a token behind it. So, to have the bank of the slot machine, there's like a small token sale, right? So, if you were to hit three uh base ETHs in a row, you'd like, you know, thousandx your money or whatever. And to have that money in the this the slot machine, there's basically a token sale at the beginning. So, this is something that I built with Scaffold ETH. I got into Cursor and Claude and I said, "Build me a commit reveal system. Uh, it needs to have a token that backs to unis swap v2. Uh, we need to do a token sale and then we need to build up a treasury. In this case, we're only building up a treasury of $15, but you need to build up a treasury of like $1,000. So, what's interesting there though is the token holders are the bank of the slot machine. So, it's sort of like the house or a co-op as the house. So, when you buy the token, you're part of the house and the slot machine has a slight house edge. And so, what happens is as people are rolling this, they end up uh putting money into that treasury. Eventually, it gets to the point where it starts buying back and burning the token. So, the token price starts to go up as people start to play it. Now, what does this have to do with X42? That's the last layer on top of this is there's an X42 endpoint. So, now your agents can go hit the X42 endpoint here. They can go to this API. And I wanted just to have a nice endpoint that I could use to test X42 out and have like an example thing. Like right now a lot of people have like fake stores and stuff because X42 is so new. So I wanted to create a fun place where your agents can go gamble with 5 cents of USDC. Again, please don't do this. This is a ruggable smart contract. Don't put money into this. But it's just an example. Uh but your agents can just run some simple code and basically they pay they sign a metatransaction that goes to the server. The server goes to a facilitator. The facilitator runs the transaction. I can run it here. Uh the facilitator pays the gas to run the transaction. So you as the user only have to have USDC. You sign a metatransaction and you get back like bar cherry or whatever and it comes right in the console. So you as the uh agent runner can have these agents that like run uh slot machine rolls for fun. I don't know. Sorry. Sorry for the over financialization and gambling. We have enough of that in crypto. But I wanted to add something fun here. So, uh, I think that's my talk. I'm Austin Griffith. Go check out Scaffold ETH. Go check out Speedrun Ethereum. Hearts, hearts. Thank you very much. — Thank you, Austin. That was awesome from Build Guild. Um, next up we have uh Med and Mine from uh Bond Credit and he'll be talking to us about verifiable credit for the agentic economy. Um, essentially how these onchain credit primitives unlock new forms of agentic driven finance. Um, let me help you hook this up. — Yeah, it's fine. I'm going to do it later. I'm going to use the Diablo just if I have time. I want to show something. — Okay. Yeah. Amazing. You — have the clicker, right? — Is it running the mic? — Uh, I think. Yeah. — Yeah. Just

### Segment 38 (200:00 - 205:00) [3:20:00]

— Okay. Amazing. — Yeah. Hi everyone. My name is Ned. Uh, I'm the CTO at Bon Credit. I've been software engineer in the last 10 years and I've been working in industry for the last eight years. — Does he have the click? — And actually, I want to ask the audience a question or two — before we even start the presentation. How many of you already interacted with Asians? — Oh, sorry. — Amazing. How many of you actually put money inside those agent? — Oh, there is maybe two or three. Okay, not so many. And I think the reason the last question is how can you trust those agent and how can you even find those agent? So at bone credit, we'll build in the credit agent the layer for the agentic economy. And before even I talk about what we do, I want to talk Oh. Oh, okay. Thank you. All right. Amazing. Yeah. Talking about credit, uh credit is something that existed since a long time from the banking system. Students get credit for financing their studies. Uh business get credit in order to grow their business and even nations get credit uh through bonds so they can finance their economy. And the agents are no different because agents also do business. They transact, they borrow, they call smart contracts, they do all those actions and they're even better than us. So they also deserve credit. So Asians are not only in crypto, they're in multiple industries. They're in dipping, they're in AWA, robotics, and they're here to stay and they're performing very well. And something I like about this industry is everything is transparent. So the latest report from RE7 showed that agent are 67% more performance than static vault. So behind me is a chart for stable coin agent. You can see that only in few months well it's the numbers August are a little bit not good but only in a few months we grow the TVL 200 200 million and that shows the adoption of stable coin agent which we actually working on but you can also notice that there is a drop on that TVL there that's not because of the agent that messed up that's actually because of stream finance so the TVL was offchain and they couldn't see and monitor the credit so the TVL dropped and they get some loss and they have to come back to the only place that didn't get lost those Asian to withdraw their phones that were saving there. So the bank of international settlements has this study where it shows that 45% of protocols on chain cannot verify fully their TVL on chain which is crazy. It means a lot of money we see it on chain but actually it's not. So capital should move with trust. Uh we've been building on arc 80004 since the beginning of it. We're biggest we one of the supporters and we actually use it to identify agent but also we interact on the reputation layer where we give score to those agent based on multiple metrics and I'm going to talk about the score that we provide in a minute and to give more trust and more credibility and also the score is verifiable and the user can come in and pay for the score using X42 you can compute a fresh score every time our score is uh locked inside the TE So it's verifiable, it's neutral for all the agents is the same score. We're not we're being completely transparent. So we want to be uh bring credit on chain fully on chain from due diligence to credit facilities to credit default. So I'm going to just show a demo about how we uh score agents. Know if that's going to work. You can tell I'm not good at this. Oh, how can I come back? All right. Moments of silence for me not being able to play the video. Jesse, maybe you can help this. Fine. Uh maybe we can pull the demo later uh after this. So thank you so much. I appreciate. Uh so basically as an agent you can come to our dashboard and register your agent. So you can give it a name, you have metadata that you can add. This is completely uh um radio on baseia since we're using ERC80004 uh pre-eployed contract. the agents once it's registered, you can authorize a party that can give you the the score. In this case, it's bond credit. So once you give the authorization uh

### Segment 39 (205:00 - 210:00) [3:25:00]

for bond credit, we will be able to pull multiple metrics. 80% of them are onchain including the performance, the technicalities, what the agent is doing, where is the agent depositing the money. We're basically doing the due diligence on that agent. And we also pull certain metrics from offchain that we're also transparent about. So we can actually compute the score and you as a user you can go and just connect to your dashboard search for a fresh uh for all the agent and compute a score for those agent and that's going to give you more trust and understanding that the agent is actually doing what it's supposed to do and what it says it's doing. So what's next? So our next release it's going to have uh an aggregation of multiple agent mostly stable agent stable coin agent including Giza they're here tonight and Zifi as well they're speaking after me and basically we're going to be able to give those a credit push we're going to push credit to them and the next phase of the product is actually allowing them to pull credit so they can have more capital. I don't know how many of you knows about Moody's but Moody's are the leader when it comes to ranking financial products and at bond we believe in the agent economy and we know it's growing and we know those agent going to need credit and somebody got to be there to be the neutral uh entity to score them and finally um before I thank our sponsors I want to say that anyone who's building on ERC4 X42 and interested in agent and credit score. Please approach us. Uh you can follow us on Twitter. We're very approachable. We're looking to build with the whole industry. We already been working with Ethereum, with IEN Cloud and Metam Mask, uh with Giza who are here tonight like our design partner and we're looking to have more users and we're going to have a beta test. So we have QR codes. So if you can approach me, I can give you some QR codes for our next release. Um, with that said, thank you everyone and I'm going to give you like 40 one minute back. Yeah. Thank you. — Metam mine. Thank you so much. — This is awesome. — Oh, if I have one minute, can I show one more thing? — Sure. Yeah. — He actually wants to show one more thing. — One more thing if it's okay with everybody. — All right. So basically I don't know if you know about um alpha arena. So we're building agent arena. So basically those are all the stable coin agent. agent that are connected to us and we're actually put money on each one of them and we're tracking them in real time. So we're going to see who which agent going to perform more. So this is something that we built for the community but also to be transparent all the metrics are pulled directly from onchain. We actually use Dune and this is something where you can go and track all the stable coin agent. This is our first uh implementation and see well I don't have internet perfect but fine. Yeah. So basically we're tracking all the agents we're giving them and we're going to give a full performance report at the end of the year about all the Asian uh stable coin agents. So thank you. There you go. Woo. Thank you, Mine. — Um, next up we have Gutier from Zyi. He'll be showcasing how you can bring low risk to uh, DeFi through agents and having more adoption and trust in that process. And ready to go. — Sure work. — Stage is yours. Oops. — Sorry. — Okay. All good. [snorts] All right. Hey, everybody. Pleasure to be here. Uh, my name is Goautier. I'm the founder at Zyi. And today I'm going to show you how agents are able to bring trust to low-risk DeFi. Um, so we've been building products in the DeFi and accountaction space since 2021. Um, and we've built a strong team. And so what are we trying to solve for who and for what exactly? Um, generating good and stable yield in DeFi should not scare you off. And this idea originated from the fact that a lot of people don't want to take time to understand how this space works, how to generate good yield and how to find good onchain opportunities. So what is Zyfi? Zyi is a non-custodial

### Segment 40 (210:00 - 215:00) [3:30:00]

smart wallet agents which basically enables you to get the best low-risk onchain yields. It automates everything and finds the best opportunities. Basically, users can customize the pools and the vaults that the agents can interact with. Agents can also rebalance your funds between defy protocols up to 12 times per day to improve the yield efficiency. So, who are we building this for? I think you've all been there um on daily basis trying to find good deals, trying to rebalance your funds from one chain to another protocol, uh claim rewards, pay for gas, interact with different front ends, right? I think we've all been there and especially people like devs and quants who create scripts or strategies to basically do these things that some users do manually. So all these things can be abstracted away. um and especially like Dows and funds who are trying to manage um their treasury in DeFi, we can also automate this. So what is ZyFi do precisely? It abstracts, automates and personalizes. The agents will claim rewards on different DeFi protocol, auto compounds, pay for the gas, and eventually will let you personalize exactly what type of risk you want to take. Unless like vaults, each user has its own safe wallet which is deployed. So basically the risk is completely isolated and everything is fully transparent. So this is for example a user who deposited $10,000 on his in one month the agent has rebalanced his fund 65 times and generated a 15% average APY. This is as much as twice as an average APY across different D5 protocols. So in numbers, uh we've moved over $1 billion, have right now 9 million TVL, and have rebalanced um funds over 50 pools and four chains, processed over 100,000 transactions. Um but one problem today is that DeFi AI has not seen a major growth yet. It's it eventually went from zero to 150 million, right? But what is missing for people to trust agents to trust automation and some basically some offchain components, right? We need more trust and credibility to reach trillions. And eventually how do we achieve this? It's with Zyi and ERC AO4. So right now users on Zifi can only access our agents by deploying a wallet, depositing funds. Our goal with ERCA4 is to make our agents accessible to anyone also um trustworthy with the validation registry. Um this I will show in the demo exactly how we implemented it but also through credibility through other agents who can give feedback on how our yield agent operates. So now it's demo time. I will show you how our product works. Um so here a user is going to deploy a safe wallet something he owns. He has a pre-made strategy that he can choose from right. Uh he can also edit that strategy the underlying pools and then the user is going to assign session keys. The session keys is going to give access to the agent to do specific things like interact with a protocol, withdraw, deposit, um have also a certain limit of transaction that the agent can do per day. So up to 12 times. This gives basically constrained rules of what the agent can do providing more security for the end user. Once the user deposited um the agents will basically look at all the yield opportunities in the market and will deposit this user's funds in the protocol which is the best um and eventually the user can come customize everything. He can customize the chain that the agent can interact with decide if the agent can auto compound the rewards of certain protocols or enable the agents to basically bridge funds between chains. Also one important aspect is like all users ask is where does the yield come from? So user can look at a specific pool the different risk parameters that our agent basically looks at. So for doing a rebalancing we're not just like chasing like looking at the highest yield right we're looking at different parameters such as like TVL safety utilization rates. Uh so a bunch of things are looked at before a rebalancing is being done. So you can see your earnings, how much you've been rebalanced per day, and then here you can see the transaction that has been done. So for example, we were on harvest 44 minutes ago. Uh agent

### Segment 41 (215:00 - 220:00) [3:35:00]

rebalanced my fund to wasabi because the APY was higher. Um and so this is where things get interesting, right? Is how do we prove that what we're doing what the agent is doing um is true basically. So for that we're generating a ZK proof which we are uploading on IPFS. Then that IPFS hash is uploaded in the validation registry. So this proves that the agent has done the execution correctly. Uh so this is already not fully live right because ERC A4 is coming out in December. But as soon as uh ERC A4 is live, we'll be able to uh to release this. Um so right now this is only available on the products but um we are working on an SDK. So if you're interested to basically um provide this agent to any wallets any centralized exchange neo bank uh we've uh created an SDK and just join our telegram group. Um reach out to us. Our whole team is here. Uh and if you want to follow us on Twitter, um here are the links. Thank you so much for listening and do let us know if you have any questions or feedbacks. Thank you. — Thank you so much, Goautier. Um next up we'll have a panel. Uh we'll be spending some time to set up the panel. So um feel free to relax for a minute or so. Heat. Awesome. So this panel is about why should we trust agents. Uh please welcome the panelists Cameron from Near Way from 1KX Nema from IEN Labs and Julian from State Capital moderated by Ben from Eliza. Welcome — right. Let's go. Firstly, thank you everyone so much for your attention so far and uh for joining us for the panel. Uh I've been looking at agents for about a year and we have been wondering when open standards would start to emerge to let agents interact with other agents and it's great to be here to talk around why should we trust agents not only in general as humans but in an open agent economy. Um by uh yeah quick show hands as well. So, who here has uh who here would let an agent manage your day-to-day tasks at the minute? Okay, that's a good number. And another question, who hopes that one day you can offload more of your tasks to agents? Okay, good. Cool. Well, look, let's dive in and start to explore the topic. Um, you know, we're going to firstly ask everybody, so what does agentic trust mean to you? And from your work, please share a bit more around what you're doing day-to-day. What is your perspective on the topic? Cameron, would you want to start? — Yeah, sure. So, um, we actually started uh building agents in about January 2024. And the problem was that, uh, you know, all these things can interact with each other. They really didn't find much utility in them. But agentic trust to me is something that I can actually run myself. Um, most people aren't going to do that. So, you do need to be able to like verify their code somehow. But the problem there is that a lot of people aren't going to want to open source all their code if they're trying to build a real business. And so uh we need to find this middle ground of like you know verifying that code inside of a trusted execution environment with hardware at the stations. And so these are some of the things that I sort of would require in order to like actually trust it but ultimately have it be open source and run it locally. Um so yeah first of all my name is Wayey going to be on this panel. Um so let me first maybe give a little bit of my background into AI. Um and for the past year I've been spending a lot of time just trying to actually understand what is going on in different verticals of AI and how performance uh is going to change now this year but also into the future. Um and my professor background is I'm an investor and so I need to understand that to understand you know what what projects invest in but through that

### Segment 42 (220:00 - 225:00) [3:40:00]

work I also realized uh look the change the transformation that will happening right now is AI is software right we're going from well- definfined well uh constrained software systems into generative and then agentic software systems that's really what AI is uh at least when you talk about right AI that's digital um you know outside of robotics and the when we talk about how can we trust agentic behavior it is fundamentally and foremost a AI problem how do we actually measure and evaluate these agentic warfalls um we we've seen coding see uh with like a parabolic rise in usage right over probably 50% of uh code um commits are coming from uh AI generated code these days in major um technology companies um whenever we have a framework to measure these um tasks then we'll see that number increase over the year after that. So we're seeing an explosion in the number of benchmarks and an increase in performance in these benchmarks that gives you a concrete score and the frontier labs are then going out to source the data to improve on those tasks and on those measurable tasks then the agents uh will improve immensely. And so for me enentic trust is can we have data sets uh for these agentic use cases and can we just iterate fast enough right to get to like fully end to end agentic workflows. — Yeah I guess um for me the way I view it is like twofold one is um in adversarial context and one is in context where the agent has like uh some kind of utility positive utility. the adversarial context is can you trust this agent to basically uh behave in the safety bounds that you think it's going to behave in. Right? So, and we still actually don't know that because you're seeing these models kind of have um behaviors that are not intended even in control environments. Right? So that's one thing when you deploy these things into the wild and you have this assumption that on top of existing blockchains which are these unstoppable computer environments if you have agents who can also be unstoppable would you want them to um act within the bounds um of some kind of adversarial bounds that you think they're operating on. The second kind of class of um use cases that I'm seeing where agentic trust is important is um if you're giving it some economic capital and you think that it's not going to um sort of mess up and um have opportunity costs for being able to deploy that capital correctly, right? So in positive kind of utility use cases, you still want to give your um assets or kind of like um exposure to agents that you think will have like the utility that they are telling you that they're going to have. — Well, I think everything was said. So — yeah, it's fair. That's fair. Um, so as a build on this question, I there's a framework from PWC where they outlined what it means uh for AI to be trustworthy. I'd like what you guys to think around which one of these are maybe the least important or that you disagree with. Um, so these six categories were human agency and oversight, fairness and non-discrimination, transparency and explan explanability, robustness and accuracy, privacy and accountability. So I think some of those are more relevant from what I'm seeing than others, but would love to hear your thoughts if any of those stand out as something that you just think isn't that relevant. — Well, I guess one of the most important aspect is um accountability. — Yeah. — Um because you need to make sure that if something bad happen that you can either recover um I mean your money or you can actually get insured. So this will be — I guess one of the most important aspect is accountability autonomous agent behaving for doing certain task but also yourself um interacting with those type of agents and making sure that whatever happens you always have the safety net for me it's one of the most important aspect of this autonomous economy — would anyone like to add to that or — yeah so I think all the popular mentioned these people reports uh are very important but maybe more of a philosophical question uh does explanability actually is actually required to have that if we know that let's say 100% of the time or even 99. 999% of the time that agent is going to do exactly what you tell it to do um right so there's I think there's benefits to kind of understanding the neuros neurosurgery kind of um the kind of neuroscience of LMS and there's a ton of progresser of like understanding what is actually going on when LM is you know generating and thinking about problems uh but at the end of the day you know if output works uh sometimes you know for example with reinforcement learning think you know um it's with long stream of stuff with different

### Segment 43 (225:00 - 230:00) [3:45:00]

languages how do you actually explain that so I think maybe that's not required if we can get to like sufficient number of nines of uh reliability — right yeah I agree would anyone else like to add or Yeah, I don't think the fairness one is necessarily that important. Um I think if it you think it's unfair, fork it and build more fair one I think is totally fair and uh yeah everything else seems more important — with capitalists I guess. Yeah, the one not like sort of counter to that but not really a counter. It's like um it depends on who is distributing these agents and if you have high distribution you actually would care or not like if the thing is like acting fairly or not. So that's where the transparency comes in right that — I think you can have transparency which counters whether it's fair or not right like you have that — I think it's accountability you know it's a well-known institution that you're interacting with then you can have more safety — but yeah three pillars of — robustness transparency and accountability — right like I guess to put it in practice it's like you know openi the agents that they have right they have distribution to like a billion user billion weekly active users. If they make like the tiniest modification to their system, like how do you hold them accountable to that, right? It's like someone else can go and make a similar agent, but they don't have the distribution, right? So, that kind of ends up mattering. — Awesome. Well, just to switch up a little bit, so you know, exploring, we're starting to explore what trust is within agents. Um, you know, humans for basically forever have trusted each other intuitively. we've, you know, exchanged services based on people that we, you know, perceive to trust or believe to trust. Um, so yeah, what are the differences that you see between trust in the machine to machine economy or agent to agent economy and the way that humans will trust agents or humans trust other agent uh sorry other humans? Camera, would you want to start? — Yeah. Um, agents trusting other agents. it just needs to perform well on that benchmark kind of what we was saying and you know I don't I'm not super optimistic about a lot of agentic benchmarks today because they're very like action specific and so if you're you need essentially to like generate benchmarks and then test against them u right now a lot of these agents are just like coding agents are like waifuss and uh I do not I think that like the main thing for agent to agent is definitely uh more benchmarks and then human to agent you got to run it yourself or verify it inside some sort of trusted hardware than a human to human. I mean, this is where like this is long- winded. Uh I'm not going to go into it, but like we have religion, we have social constructs, we have these things that make us uh trust each other more. There was this famous quote like, "If you don't believe in God, I don't believe in you. " It's like a lending market like for people to lend back in the day. And uh yeah, I think that that's not necessarily going to replicate to the agent world. — Yeah. So, um we're at Ethereum conference, but I'm I'm mostly going to be talking about the actual AI parts. Uh so apologies but here's how I think about um how to replicate some of the why humans can trust other humans and other businesses uh the same type of mechanism into the AI world Asian world. So first for and foremost I think you know how many of you have used Google maps where like equivalent service uh within the past day right and saw the reviews you know right like you see a place like four stars and you know maybe not like five stars I'm going to go there right that's a review kind of identity system and I think we're seeing the equivalent of that right with uh 8004 um so that that's one component right in a lot of real world scenarios humans trust businesses through review processes of you know when other users have uh liter reviews uh and then there's potential like actually regulation can help here. So when you go to a restaurant any restaurant in a uh you know modern country there's regulations on how that restaurant can operate right there you have you have food quality standards you have safety standards and so when you whenever you go there you assume that they have a license to operate which means they have pass inspection and so for certain AI use cases uh for example AF for law or AI for medical use potentially uh we actually have regulation on what type of system that can be deployed and Then uh regulation is really maybe the really the bottom line, right? And how many of you um the third category, how many of you uh have um consciously chosen not to fly Boeing recently? Raise your hands um or like you know thought twice about like getting on a Boeing plane — always, — right? Uh and you know because yeah all the recent quality control issues. Um right. So it's also yeah maybe they have pass regulation they're under FA investigation but it's the actual performance right like how many nines can you get to and what is the track record uh um you know this is not what other people are saying but it's actual track record right so that's also tying to what I was saying before if you really get to sufficient number of nines

### Segment 44 (230:00 - 235:00) [3:50:00]

you have this brand that people trust uh you know I think the same thing can happen with agents and uh you know for example maybe you'll trust Google Gemini more than openai maybe you'll trust near agent more than all the other ones U but you know those are the kind of three categories I think about. Uh so right reputation systems um and you know potential regulation to get that um back stop and then like what yeah what is actual performance what is the number of nines you can provide here on reliability. Yeah, the way I see this um is mostly rooted in the speed at which reputation kind of like evolves um among humans and like among agents, but then also like between these two groups of entities like the way that you think about another like business or another human, right? Like your um perception of that counterpart um is a function of like real time like if you perform an action, right? Then I can decide my what I think of you as a result of you doing that action, right? But that's like based on physics. It's like how long that takes, right? Whereas like for agents like the speed at which like they can interact and they can do stuff is like dramatically higher than like what we what we're used to as humans. So what I think might be interesting to see is when this agentic economy kind of gets off the ground and agentic systems have this identity have this notion of identity and reputation. um can we like reliably keep up and actually trust that like a certain agent has that you know reputation like did it deserve it right it's like because it evolves faster than we think it does — well pretty simple u human trust is emotional machine trust is statistical meaning that you will probably see a lot of interaction between agents that feel extremely dangerous or insecure for you but because those agents they have adverse professorial gaming uh mechanism in place that they can hedge themsel leverage or know that they can win. They will make decision that is not necessarily the decision that you will have made yourself but it can serve their interest. So it will be very interesting to see Asians interacting between each other in a way that is completely different from what decision you will have made yourself. And this will be a complete new economy where you will have to measure the risk but not necessarily understanding the short-term implication but more like the long-term um goal or milestone that the agent is trying to accomplish. — That's very interesting. Um and like zooming out a little bit the you know it was said this morning Marco and Dev have been working on 8004 X42 we're building these open standards for the agentic web. Right. So you just touched on some of the risks there as well. Like history has shown us that perhaps when we haven't got these standards right, things have perhaps been closed or systems have been, you know, say misused or have perverse incentives introduced. Um, you know, what do you think could go wrong if this open agent economy doesn't have these standards or standards for trust embedded kind of now before we realize what it is? And Julian, I don't know if you want to start this time. Uh well something I've been thinking about it for quite some time now. We're building a decentralized economy blockchain being able to have autonomous system. So in this case agent being able to interact between each other. We also been um I mean as tech capital we've built a facilitator and the lead developers here that integrate uh for agents to be able to make payments between each other and then also the trust layer. What I'm trying to understand is um if we have this system that can like we cannot stop that can interact between each other and we have autonomous agents like sophisticated AI. I mean a system like that is really nice and really helpful if it served humans or in control by humans but when you have such a system that is full control of AI then AI can actually do things that you don't even have knowledge of because you have verifiable compute or whatever like encryption fully ZK mechanism in place then those system can actually call you between each other do things that you have no understanding go as fast as speed of light to your own uh time frame. So I feel like for example like I've seen the manifesto from Vitalik but well first I think the manifesto was extremely dry. I mean okay but I was a bit disappointed by what he the article but secondly I don't think we actually spend much time understanding what how the system will be used in a fully decentralized autonomous and extremely intelligent world are they going to serve humans or their own interest and how can you stop them and I feel maybe as an ecosystem

### Segment 45 (235:00 - 240:00) [3:55:00]

I'll be the first one who will be trying to bring more centralization to system in order to make sure we can stop them if something goes wrong. Um I think we haven't thought much about it because now we're trying to push for privacy for autonomous system for super intelligence system but we not having thought about okay what happen if something goes wrong — you blow up the data centers. — No — you what No, because the problem is like if the agent has equity in a fusion nuclear plant plus data center control the entire supply chain, you cannot really stop anything and also you don't actually know what's going on. I for example like today I wonder if statistically speaking or if there's a prediction market that can predict are we do we have AGI up and running or the question is do you believe you have SI super intelligence up and running the question is like for a super intelligence machine they will be not incentivized to reveal themsel because they understand that they can be stopped right now so maybe you have it but you don't know it a super intelligent system will do that will still be able to interact between and do things but not reveal himself because he knows that right now is not the right moment let's wait a little bit more have sufficient u acquisition to technology for me to be able to protect myself if they realize that I exist — so great — counterpoint to that like the more time I spend actually understanding like application of AI systems and uh into these inter enter enterprise settings and what when they succeed when they fail and if you look at the benchmarks yes there are risks to potentially AI fooling their benchmarks you know there just recently if you look at the Gemini 3 safety reports right there's uh basically like the model realizing is in a test environment and switching behavior accordingly because it realized okay I'm not in a real setting uh you know there's stuff like that the you know the model is getting um increasing sophisticated But it seems that the consensus is we're very far away from AGI and it's all about improving uh performance of agentic workflows on specific domains and we need better data. — I'm not sure about that. I mean OpenAI spent two years trying to remove the double dash and it's still not removed. So — yeah. So we're very far away from — safe. I mean — AJ people have been trying to say okay can you please remove the double dash? dash when you're writing content? And it just announced couple days ago, Sam announced that they made it. But try it out. The new version is still — Oh, yeah. — The double dash on going. So, if you can remove a double dash, how can you make sure it's safe? — It's you can't remove double dash. How can you make sure that's AGI is, you know? — Yeah, that's what I'm saying. How can you make sure it's not actually a GI? — So, this is Near's been Ilia co-rote attention is all you need. I'm assuming a lot of you guys have heard that before, but um they've been thinking about this for a long time and their solution is they want to formally verify the entire internet. And if it's going to be, you know, inherently unsafe and agents will exploit every vulnerability on the internet, well, you're going to have to formally verify everything. And so the goal, this is what Alex uh the other co-founders been working on for honestly since Near started uh is to train ma like get mathematicians to create data sets to learn lean to generate a model that can then write lean code which is for those who don't know what formal verification is. It's essentially a way to verify that the code is operating the way that it's supposed to. And there's all these ways to like ver like write mathematical proofs and then prove them. And uh there's not enough people in this world who know lean. And so the whole point is build a model that can do this and then regenerate the entire web and pretty much patch all bugs because AI will exploit every vulnerability. And I actually don't think regulation is going to play a huge role because I live in San Francisco. I know what it's like when uh the system gets overwhelmed with crime. And uh you can't actually call the police because they won't come and uh I see that's going to happen with like the legal system when agents exploit everything. — Yeah. I guess um to go back to the point of like the safety right and how these models are thinking um not sure how familiar people are like with the internals of like the AI systems but like sometimes when you look at the chain of thought you have this like language that some models have called neural right where like you have no idea what it's actually trying to say um and if you ask it a question in English it thinks in ch Mandarin and it like responds back to you in English so it's like we still have we're still very far in figuring out how these things actually work So I do think there is merit in um what Julian was saying. Um and I do think it is like relatively far but I do think if we don't do something about it is pretty

### Segment 46 (240:00 - 245:00) [4:00:00]

inevitable because whatever standards come up with 8004 x42 whatever right if they're like to some extent unstoppable right and they if they can deploy like code that we don't know how to interpret as humans right um you can wreak havoc very easily — can maybe I actually address the original question which is on open standards um it so here's how I think about the the two paths, right? So if when you say decentralized AI, what are we actually decentralizing? I think the most important thing to think about is actually uh proprietary data, confidential information and kind of business knowledge is siloed in different organizations, right? and and this also applied for individuals and it's the decentralized uh system of proprietary information that demands decentralized AI meaning the AI systems that individuals use should align with their intents and the AI system that the business use will have access to their propritor information and you know all the past business records and act on their best interest and instead of open AI having an AGI acting on everyone's behalf you should really have a agents that represent the best interest of each individual and companies and that's where we need open standards like the internet right we don't want Facebook uh you know Tik Tok centralized platform harvesting data we want email we want HTTPS and I think that's what why all these open standards efforts uh is going the right direction — yeah I agree and I think uh there will be more open standards needed to make sure that things are perhaps contained in the right way and the regular regulations aren't going to do this right. Um, yeah, thank you so much for those answers. That was a great dive in. Um, so looking a little bit ahead, you know, why should we trust agents? Uh, we must all believe there's some premise in this technology and it can better humankind in some ways. Um, how do you think the world looks if agents are successful, they're trusted, we're redefining the economy, we're, you know, we're adopted by people on mass. Um, and is there anything that particularly excites you personally or anything that you think at the moment is maybe a lot of hot air and isn't the right path? Um, yeah. Julian, again, would you like to start? — Well, it's uh it's obvious that it will unlock uh a tremendous amount of u value in the world. So 24 hours, 7 days a week will be uh full um capacity. So you'll be able to have increase of productivity, efficiency, transparency. Um all the repetitive task that you do on a daily basis will be removed from your plate. So humans will become more focused on what's the goal the visions uh of things that can help themsel or can increase or produce more value. Um so for example I believe like in the near coming I mean the coming future you will not see apps on your uh potentially less apps actually on your phone because it will be intuition coding. So for example you'll be thinking about something and then automatically these things will be created on your phone just for you just for your feet uh based on your needs based on your emails based on your data information. The reason for that is because the compute will be so cheap that it will be cheaper than actually make something happens on the fly than actually hosting that app somewhere. Think about warehouse in AWS or whatever. It's really extremely costly to store items. So they prefer sell them at loss. It will be the same type of things. So you'll be able to build things on the fly, come up with um you will not have agency building apps for you. You will be the one building everything will be from you. You'll be your own economy, your own enterprise. So we unlock a tremendous amount of value. — Interesting. — Anyone else want to pick that up? — I mean, yeah, my take on this is like we don't really have a choice. we have to trust them because whether it's like defensively or offensively like if you if like you know they're going to it's one it's probably the most sophisticated technology ever to be developed right so if you can use it for malicious reasons you want to counteract that malicious activity or if you choose to like not deploy it and like not kind of lean into the positive aspects of AI as well like you know economic flourishing and so on and so forth entertainment whatever it's like you're just like limiting yourself from a better human experience as So — what I look forward to the most is this shift from uh human decision making being the bottleneck in most kind of consumer applications right so there's

### Segment 47 (245:00 - 250:00) [4:05:00]

this theory that if you present too many choices then you know the consumer would not actually choose anything because there's too many choices uh but that's no longer true when you have an agent that you trust and I think that shift from uh make less way more choices for the consumer and you have okay you know just take the subscription and you're done you know uh and everything is subscription these days move from that to my agent making decisions for me and making micro payments based on what I actually consume and what I actually like personally. uh and you know how many of you booked um spend more than an hour booking a trip to Buenos Iris right like okay I'm surprised not that many hands came up you know it's it's a hugely complex you know to like book flights book hotels right figure out where the conference venue is um you know all that will go away in a if we do execute a aentic economy right and um yeah I think we can get there I think we'll get there um I think that far that war is not far away — yeah I'm super aligned with what Julian was saying about not being apps. And I'll walk you through like exactly how we planned to do this because uh first off like again we built an agent framework a long time ago realized there was not many useful agents being built. So we pivoted part of the company just to build useful agents for enterprise where they actually open up their they pay us to you know give us their data to build useful agents for them that we can later generalize but I don't think the initial version of this is be agentic. Second piece is we need to ensure it's everything's private. So that's where the private inference comes in first. Uh we talked check out uh Brave's recent Brave actually just announced yesterday uh they will be powering a lot of their LEO agent. For those who don't know there's actually a Gentic browser within Brave uh powered by LEO and they'll be using Near AI uh inference for to verify privacy. So that's like pretty cool. 100 million people uh will get access to that eventually. Um, and then the next part is like a private portable memory because you're going to need to move your context from agent to agent or like you know system to system. And then after we have private portable memory, we'll generalize these agents that we've built uh for enterprise and other places. So then you have your agent that verifiably belongs to you that can then communicate with everything else. And that's where these open standards come in. But we have a long way to go for that. Uh people need to completely abstract the on-ramping process away. people like buying credits like for any AI app that you use today needs to be the same but you need to be on-ramping stable coins into an agent wallet that can then transact on any chain and that's where the whole like agentic commerce piece comes in. So there's just like this giant web of stuff that needs to happen but there are these fundamental pieces that are kind of missing today specifically around insurance security um you know privacy is getting there but uh after that yeah our agents will just sort of communicate ideally live on device uh call larger more powerful models via private inference APIs and uh yeah you won't have to work again — and on that bomb show thank you so much for the panelists for joining Uh can everyone give them a round of applause? That's the end of the session today. — Thanks guys. — Yeah. Thank you so much. Heat. Awesome. And next up we have Alex from Renit Labs. He will be presenting to us um insider agent trustless market maker intelligence. essentially putting market maker trading into TE to make sure that they're not front running, they're not doing me, they're not doing bad stuff. Um, and also having agent coordination in between. So, let's welcome Alex. Awesome. — Let me do the Sorry. Oh, sorry. Okay. Gotcha. — Yeah. Clicker. — Oh, cheers. — Yeah. And now it work. Okay, perfect. Hey guys, I'm Alex. Nice to meet you. I'm co-founder of RenaLabs and we are building the verifiable you know onchain reputation layer for traders

### Segment 48 (250:00 - 255:00) [4:10:00]

because we notice that in crypto we see so many shady deals behind our back and many foundations and founders they complain about it. So that's why we're here to solve it. We want to come up with the trader reputation protocols which means you know as a you can see the trader histories you can see this you know address histories if it's front runninging if it's a shitty deal and the first case we want to focus is market makers um there are four hurdles while we are doing that the first one is opaque performance so imagine uh when the market maker tell you they don't dump in the token but what about they get they don't give you the API how can you value it or they give you the added data you don't know the Next one is like it's too fragmented because there maybe there's so many places they're trading CEX DAX or even some D5 protocols you never know and next one is no standards because even though we already help more than 10 foundations to do that but there's no unified metrics working on that so you don't know which metrics is best one for them even though you do how to balance different metrics and next one is some foundations they may hire some advisor to do it but how do you know the advisor they don't take the bribes. That's why we are, you know, we are willing to use the agent to fill the gap. And this is our solution. We're more um based on our first product inside of the cache. We help foundations monitor the market makers without exposing the raw data. And we come up with the trader reputation protocol which is you know integrated with the ERC 8004 because in this protocols you can see this traders or this trading address the past histories and how's the performance and what's is a ref flag or not and the next one is all the data you submit in our platforms it's protected by the TE which means you don't need to worry about your privacy will be um damped or others will see it and in our demo We will have two parts the analyst and the rating system. Uh for the rating system we're already pre-trained because we have um private data provided by more than 20 market makers and also 10 foundations. But for the analyst it's open for everyone. You can give the prompt or based on different models just by yourself and you can use the rating system to read it to see if your uh analyst is good or not. And this is how it works. First you just register the agent and then submit the data and all the data is submitted like I mentioned is protected by the TE and then you can use the agent to generate the reports and our rating system can give a rate then you can upload the onchain attestations to make sure that all the performance will be stored in the IPFS and the last one you can report the reputation profile which is same by everyone so everyone can check it onchain and this is our future milestone. The first phase will finished like the market maker uh monitor inside of the cash and the next two we will focus on the profile you know the professional traders quan traders and also the retail traders they can use the same algorithm to protect themselves in the last we can come up with the onchain reputation network which means all the data you are using can beneficial for others so let's do the demo because I recorded it last night because the internet is not that good okay uh let can do it. We have two functions agents and — so you can see there's um know two functions create agents and there are two rows first is the analyst the second one is the raers the analyst which means you just use a model to just talk to him and they can follow your instructions and readers just rate the analyst the per um the um answers and you can see in the later and just input like MF George Here you can judge different market makers descriptions. Uh in this demo we are using cloud anthropobics and choose the um 4. 5 upload API key and uh I already you know printed a prompt so I just input a long prompt. Okay. Now it's already you know created the image just too and when you click you can deploy it to the blockchain but we can do it later. We first go back to uh you know use it. And this one I already tested last night, but this one we just use a new one. And let's use maybe winter meal.

### Segment 49 (255:00 - 260:00) [4:15:00]

And then the agent will working and it won't it will take some time sometimes. I think it's probably like less than 10 seconds. Let's wait a bit. Okay. As you can see, it gives a winter meal, you know, a rating. But because at this point, we didn't give us um any know private data. So you cannot see very detailed but we can use the you know I already created the rating you know system so you can just read it by the insider agent you will give you know evaluation you can see the score is not high because we didn't give a very detailed data it's based on the data you give it to him then you can close it and back it then you can upload this um performance rating to the blockch again. So this 50 deploy and then waiting for a little bit after it was uploaded. You can check on chain and also see the histories of how's this agent and what's the review. If more people use it, there will be more reviews which means this agent is more reliable and it can be used for more people and the first visa. Yeah. And we just reveal by the ether scan. Okay, then it works. I guess that's all for the demo. And uh if you're interested in love to connect with me and also all the uh trader you know protocols and let's build this you know crypto with more transparency. Thank you so much. Thank you Alex. Um and next up we have another Alex. Alex Coopermanman from Olas. Feel free to come up. And so he will be giving a talk on bringing visibility to agents from how to make agents visible, transparent, and indexable across the ecosystem. Uh let's give it up to Alex. Let me help you. — There's a clicker. — What? — Yeah. Just the green one. Just the green — to do — to go to the next slide. Oh, — okay. — Yeah. — Start mirroring. — Yeah. Yeah, I can quit for that one. Yeah, — sorry. — Okay, sounds good. All right. Hey everyone, my name is Alex. I'm a senior smart contract developer at Valerie and the founding member of Folis. And today I'm going to talk about bringing visibility to agents. I'm going to touch base on two key points. One is how Olas uh pioneered AI agent registries in uh 2022 and is now ERC A004 proofing it. And uh second um how we plan on making discoverability of agents easier. Um, Olas has been building agents um, since 2021. Um, and uh, gained a lot of experience there, a lot of traction. So we're very excited um to see that the um concept is coming mainstream because back at that age um even the LLMs were not um on the main stage yet. Um so uh the track the history of all can be looked up at all. SL timeline or by um scanning this QR code. Um throughout the presentation I'm going to show what we've done so far and how this correlates with ERC 8004. Since day one um we've uh had a vision of creating a platform that uh allows for true co ownership of AI. As most encrypters say on your keys, own your coins. We push that to AI agents with um the uh um phrase that uh on your weights on your brain that translates to uh your AI agent owns weights of your models. Olas has created the largest AI agent economies encrypted to date with more than two and a half thousand agents

### Segment 50 (260:00 - 265:00) [4:20:00]

being registered. uh hundreds of agents uh daily active agents that serve different agent econ economies with more than 11 and a half million of transactions and keep on growing rapidly. Um among which more than 9 million transactions are agent-to- agent interactions. We have deployed the first AI agent registry on Ethereum and now serving nine chains. the protocol being largely um chain agnostic. Um everybody like literally any account is able to get their agents registered and uh owned at our um marketplace which is called marketplace. all. network. You can see there are lots of agents floating around and each of them is uh represented by the set of configuration parameters etc that shows whether the um agent is active and doing its tasks or still in the setup mode and also the full transparency about the transactions that agents do. They also present uh on the website. So the trustless agent registry gets us to the architecture that allows for a full um agent life cycle management with a representation of a agent um by two um simple but at the same time very effective entry points NFT and a smart account attached to it. There are two types of OS agents. Sovereign agents, those that are more lightweight and can be run on your laptop or in a cluster. Uh and decentralized agents um those with basically ST security network of nodes. Um but the philosophy is the same as for soaring agents. On the right hand side of the slide you can see the full FSM of agents going from registry to being we call it deployed become operational. Um the workflow is pretty sophisticated. So we've created the um framework called open autonomy. If anybody is interested please see me after the presentation. Um we're going to talk about how to uh construct your agent and bring it on chain. That's very exciting. Some of the sovereign agents have become canonical in our ecosystem. For example, Baby Den trades autonomous autonomously in DeFi. Um, as for the decentralized agent, there is mech agent for example that serves agent interaction via AI tool marketplace. Agents are not going to come because they're already here, right? and they're just going to grow in numbers and u um qualities as at the same time as more registers might come onchain they could be constrained to just a single registry right that's why it's great to see that uh there is a need for emerging standard which is CRC 8004 we're also very happy to have contributed to uh the standard like for example um identity registry design follows NFTs as well as the means of um ownership and trading management tools for AI agents. Validation registry and reputation registry bring new functionalities. Uh very great to see new uh new tools come on chain. As for our registry, we've started way earlier, but the composibility of our onchain stack um in a core peripheral logic allows us to uh quickly plug all of our agents um to be discoverable by your 0004 registry via the identity register bridger. So we basically um link um all agent NFTs with identity registries once and uh the all the information is automatically synced uh without any um you know interruption of any of the contracts. There's been a lot of standards coming to AI agents lately and uh the agentic world right now is so much different uh compared to what it was in 2021 and the natural cost to tackle this challenge would be to try and make it easier to discover agents across different standards. That is why we would like to invite all the interested contributors that are interested to work on agent scan with us um in order to

### Segment 51 (265:00 - 270:00) [4:25:00]

facilitate the discovery of agents because we know every agent deserves a place and uh no one has to be lost. So please join us in building an open-source discovery layer that's cross standard chain agnostic and uh that's um going to uh create a great course for um all of the agents with their standards, current parameters, settings, etc. We've all probably seen what's happening in the crypto market today. We're not sure if trillions are coming to crypto as coins, but we're that sure know that trillions of agent transactions are coming and super very soon. Thank you so much. Thanks, man. Hello everybody. I'm going to replace uh Jessie for the next hour so she can have a bit of a rest. My name is Marel Boli. I just joined recently the Ethereum Foundation, the decentralized AI team and I'm going yeah to follow Jesse work for the next hour and uh obviously there is a problem immediately because the next speaker is not here. So we're going to have a rest of five minutes. I can have the his speech. I go ahead or Okay. So we're going to have a coffee break. Um and uh we see you in uh 5 minutes. — Okay, just stay here. No coffees also because they don't have Italian coffee. So thank you — Hello. Okay, the break is over. So, we're starting right now. I will introduce Andre Samra, CTO of Warden Protocol and he's going to talk about why UX matters more than the infra. Thank you very much for making it this far, I suppose. Um, so I'm Andre. I'm with Warden. um to give you that quick minute pitch. Uh we're basically Oh, and we're not displaying anything. Can anybody guess it's a break? — Uh there it is. Okay, cool. I think it's just over there. Some problems over — Yeah. Yes, now it is cool. So, uh, thank you for making it this far. Um, so Warden is trying to place itself as basically the launch pad for agent developers. So today I'm going to talk to you a bit about why UX matters most more than infra and why this is important also for developers. So our this is our premise basically — his moderator — ask you all at this point if I was to ask what is the de facto wallet for Ethereum — can anybody answer what is the wallet — okay that's it — cool okay so for a lot of new users this is a very difficult question in fact because not only do they have to

### Segment 52 (270:00 - 275:00) [4:30:00]

understand what is the best wallet for this particular network, but also they're going to find out there's all these other networks out there that each come with its own wallet. Managing that kind of complexity is completely ridiculous for the regular user. So, of course, we'll have the power users who are pretty fine with that. Um, but the premise that we're starting from is actually that we can use natural language to address this complexity without having to understand it. And to do that, what we're starting from is agents that act as the interface between the users and the actual outcomes that they're hoping to have. We don't use or we're not trying to push for the concept of an agent that takes care of everything that you need to do. they're there just to translate intent into actionable things that happen. So the way we did that was we wanted to abstract complexity but also give you the safety rails. So when you go and you do a transaction, you don't have to worry about, you know, what happens in the background like am I going to trust this app? website? Am I going to trust this platform? We're taking away that complexity. And we believe that better UX leads to increased adoption. And in the six months that since we've launched, we've seen huge adoption rate. So we're past 16 million users at this point. Um people that really think UX deserves better in uh crypto. Now this is just a quick screenshot of how this works in practice. So we have a mix of UI elements and natural language in the interface that only complements the missing bits where you have to have the hand over what happens right so let's say I want to swap here right I want to do a transaction you know doing this with metam mask means popups coming up understanding what goes there what are what's the hashed message that is being displayed what is the information that I have to understand here it's very easy for any user to understand what they have to do right they just need to you know double check that everything looks cool and they click a button and the transaction goes through now we get to the second part which is that if we fix the UX we'll have adoption right we're going to grow the user base and if we have the users that means we're going to convince developers that they should build with us right and to do that we basically need to show the developers that if they launch their agents with us they have instant access to distribution right the loop is very simple you build or let's say the pitch right for developers you build your agent You launch on Warden permissionlessly and then you get paid on day one. You don't need to care about anything else, right? You're a developer. The only thing you like to do is build stuff, right? You write code. That's your purpose. You don't do marketing, right? You don't do user acquisition. You just want to build stuff, right? So, that's what we're going for, right? We want to give you the peace of mind to build the things you want to build and we take care of all the rest. Now obviously this is not going to work without standards because you know we don't want to enforce a specific framework or a specific tool set on the developers. We want them to build using their own frameworks in whatever tools they want. We just need to use standards to communicate with their agents. So that's why 8004 is so important A2A and X42, right? Because it gives us this crosschain reach. So we're fixing this coldstar problem for developers and to incentivize them. Uh we've launched actually this week this $1 million program in incentives where the first month, you know, the top 10 agents that are being uh released with us, they're going to get $10,000 each um in incentives. So hopefully that's going to help them. Um it's going to help convince them. Right now I want to do a demo. Um I have two minutes. Let's see if this actually works. If you are in this business and you don't have an appetite for risk, then what are you doing here? [snorts] Um so

### Segment 53 (275:00 - 280:00) [4:35:00]

this is Warden Studio. This is the platform where developers go to um to register their agents, right? So let's see how long this takes. So what I as a developer the only information I need to provide in fact is the API URL of my agent and optionally in case the agent doesn't use X42 which is the case for Langraph agents I cannot do much. I we can already see here that there is some information that got prefilled. This is coming actually from A to A. So the agent card information. Uh I'm just going to make it prettier. Um weather agent. So dev connect. Okay. This is uh an information and it has a an image that we're going to put here really quick. And then I'm going to set say this is I want to charge one cent for my agent. And what happens next is that we are going to register this agent using A004. So we're giving it an identity. We're giving it a wallet. Um so it's all on chain and then we are going to store a lot of that metadata for the agent in fact um on chain and offchain for easier discoverability in the app and then once all of that information is done um the registration is complete. So hopefully that won't take too long because we only have 30 seconds but should be enough. Um we're just going to wait for this process to complete. So in the interim um what we also want to do for developers is to help them launch their agent, right? Oh, cool. This is done. Let me quickly go through that. So now, oops, here we go. So we can see the agent has been created. It has uh an agent address, the A004 identity address. And now what's left is to go and use it. So in the warden uh community section of the hub, we're going to do a quick refresh here and we should see that there we go. We have this new agent that is now available. So clicking on it is basically going to just allow me to use it instantly. Right on day one, zero validation process. It's instantly available. So I can ask what is the weather in Wenos IRS. Now what's going to happen is that um the app is going to withhel payment so that it basically charges me first and then it waits for the agent to finish work and then that's it. It made the payment and you know you get quick refresh again and you can see here that you've already started making money. See you have earned 0. 01 in fees. It only took a few minutes. Thank you all. Thank you very much. So it looks like building and launching stuff on top of the 804 X4JU is quite easy right now. So I would like to invite the next speaker YQ founder of Alt layer. He has been also really involved in the 804 builders community. Uh he developed this 804 scan. So please the floor is yours. Welcome to the stage. Thank you. — I think it's this or that. Yeah. [clears throat] Um, hello everyone. Uh, glad to be here and to share a little bit more on like sort of uh what we've been building for the past few weeks and uh yeah. Oh I can't go back. Give me a second. Okay. Yeah. So, um as I just mentioned right YQ here, founder of all. So, basically we are building a lot of these infrastructure solutions for blockchains especially Ethereum and uh I just want to do a quick quiz. Uh have you been really following like sort of X42 and AO4? Uh any of you read any of these blog posts before? Oh great, great.

### Segment 54 (280:00 - 285:00) [4:40:00]

Thank you. So um yeah so uh recently I've been writing a lot of articles uh to explain like what's uh X42 and also AO4 and beyond that I also build a bunch of uh these kind of tools public good tools uh to help people to index all the agents around this AO4 and of course like in the previous talks I think the speakers already share a lot about like why we want to have this agent and in general what's the agentic economy right and based on some lot of this analysis either from Mckenzie or from the other ones in the future we may have like over $4 trillion dollars uh this kind of value around these agents and um right now based on all the ex uh existing crypto like sort of agent stacks as we can see basically we have X42 to cover the payments and also this AO4 to cover this kind of identity and also this reputation layer and beyond that we also want to have this AO4 scan to really help people to discover all the different agents in the space. So that's why if you look at this diagram right the thing is like in the future if we really want to create like sort of this onchain agent the organic this kind of uh life cycle will be like first you register this kind of your agent identity on this AO4 and after that it will help you to mint the NFT and beyond that uh you also need to fill in the points and also this capability all the different fields and beyond that you can also immediately index by this AO4 scan and after that of course you can leverage this X4 to your payment. So in that case it will be gas list and also payer use payment and beyond that um AO4 also provide you this kind of feedback and the reputation system. So later on you can get all the feedback from your users for your agent. So for the X42, I believe Kevin also Eric mentioned a lot about why we need X42, right? Compared to web two, it basically has much less cost um uh regarding like sort of the either the infrastructure and also the overall like sort of charges against the users. And beyond that, if you already build something on this X42 v1, you will see that there are a lot of pro and cons. Of course, the best uh benefit we have is really the gas list and uh if we are using this coinbased facilitator, basically no one need to pay the gas and the coinbase will cover that kind of cost and beyond that there are a lot of cool stuff about this XO2 and as you can see right this um so basically we can easily convert any sort of this uh payment URL um all this kind of service URL under this pay wall of this XO2 immediately you can convert any of your service into a service like everyone need to pay. Uh, of course there are some cons like as limitation, right? Uh, there's no atomic settlement and also sometimes it doesn't cover this multi-token like sort of a uh standards. Um that's why like um a few like sort of days ago I proposed this extension to X42 basically help to really fulfill a bunch of these challenges we have like for example atomic uh settlement layer and also adding hooks and multi token support. Um I have a quick video and uh like exactly yesterday we already launched this kind of demo for you to see how you use this extension of X42 to quickly launch uh like sort of your service uh under this X42X and immediately you can have all the benefits for example multi-token support and at the same time you can this uh standard and also gateway can immediately give you tons of dollars uh transaction per second. These are super high throughput. Um and via this website later I will share is basically you can visit x42x. app and uh you can basically uh launch your um gateway at the same time you can also manage your uh service urls and uh you can config delete and also uh further um send this kind of payment URL to the other users and the other friends. Uh and beyond that um as you know right we also have a bunch of this coverage this uh AO4 and uh especially right now with a lot of agents we built in uh web two or crypto we want like sort of to have a um authentic list of all the agents so we can easily query the other agents and leverage the other agents to achieve some of complicated tasks. So that's why we need this AO4. Uh, of course there are lot of um details [clears throat] already covered by David and also Marco. Um, basically it can provide the data ownership, portability, persistence and

### Segment 55 (285:00 - 290:00) [4:45:00]

also transparency and censorship. In that case like compared to MCP or A2A, right? this AO4 can provide all these very nice decentralization and uh anti-ensorship uh features uh for all the agents and right now if you directly visit AO4can. io Um, it automatically index all the agents uh register on this AO4 contracts and uh it basically can save a lot of time if you uh want to quickly search the other agents at the same time if you really want to uh build on top of these agents. This kind of scan really give you a faster like sort of the track for you to do it. Um right now we already have the basic feature like explorer and also uh this kind of analytics and later on we will further add this kind of feedback system reputation system and exactly before my talk we already enable this feedback system. Right now all the users can give feedback to the agents and at the same time agent can manage all the feedback from the users. Um [clears throat] as I mentioned at the beginning for this uh talk right the things like right now we already have a bunch of tools to build agents and at the same time we want to further help developers to have better developer experience. So that's why with this new scan we want to help this all the developer quickly search like sort of their agents and beyond that we uh with all the different futures and also the API endpoints in the future uh the develop developers can also use the CRI to quickly interop with scan and also this AO for smart contracts. Um and uh this is sort of like one of the use case and for the life cycle. Uh so you register your agent on this uh AO4 uh either via scan or via the other tools and then immediately will be indexed by the A4 scan and after that the other clients and user can discover this agent can start to use and they can further use X4 to do the payment and then later based on like sort of their user experience they can give you the feedback and set up the reputation for your agents. Um and in general as I mentioned right this X42 AO4 and also AO4 scan we want to have a complete stack to really cover everything about this kind of crypto agents. Uh we already have the st uh you can just visit a website um and we also have this quick demo and probably just a few seconds u but in general you can visit right now it's a4scan. io and uh you can use it try it especially with the new feedback feature. Uh it's pretty cool. We also have the inbox. If sometimes a message you haven't received, you can just check out your inuh inbox and then to operate later on. Um and also they are providing the whole suite of X42 uh this kind of services factor gateway and also some other services. Just feel free reach out to me if you are super interested in X42 and AO4 and uh if you want to read articles and just feel free to follow my Twitter. Uh thank you. — Thank you. Brilliant presentation. So next, so today we have uh ways to register agents, find agents. What is missing in my opinion are nice agents. So agents that can do stuff. For example, after that connect, I'm going to Patagonia. I would like an agent that can book my flight, do everything for me. So we need people that bring value into the community with their agents they're doing. Marco de Rosi few days ago introduced me to a guy that is doing this and this guy is the CEO of Giza. So I will introduce him. Welcome. The stage is yours. Thank you. All right. Can everybody hear me? I was in the back and I couldn't really hear a lot. Um my name is Ranch. I'm the CEO of Giza and hopefully you guys all had an amazing week. Uh this has been an incredible event. At Giza, we focus on the biggest product market fit of blockchain technology which is decentralized finance and we focus on preparing it for its next evolution where the capital itself becomes autonomous. Before diving into everything, I would like to ask all of you why are you here? or why are we here? Why did you travel countless of hours from where you were and came to Bonosirus? Besides the amazing food, besides the amazing city and besides the incredible scenes, we are here because we are innovators. like to innovate and push the boundaries to bring adoption to bring growth to Ethereum ecosystem. And that growth and that adoption goes

### Segment 56 (290:00 - 295:00) [4:50:00]

through simplicity, goes through accessibility. It goes through democratizing access to finance, which is the biggest part of blockchain systems. And democratization of finance doesn't mean everybody in this room gets a wallet. It an inflationary token in their wallet. It means that everybody in this room will have access to financial opportunities that they can leverage without any limitation. And this is what we do at Giza. The reason why we take on this challenge at Giza is because currently our financial workflows are quite complex. And I can tell you how complex they are, but I can show you better. This is a snapshot of the stable coin ecosystem, a part of it today. By the time you go to another event this tonight, there will be five more tokens added to this. This is simply put inhumane. No person can adjust this complexity. Currently, all treasuries, all funds that we talk to are trying to run a stable coin hedge fund using consumer tooling. It's time to upgrade. It's time to upgrade because this is the only way we can bring more people to Ethereum. We want to take all of these things and turn it into this. An agent that is acting on your behalf with the policies that you set for it. Giza agents remove countless of hours of operational financial overhead. We provide you with all of the risk management tools that you can to set your policies. We run your capital 24/7 on behalf of you because your capital deserves it. And in return the users that we have are getting 2x higher risk adjusted yield than they would by using a static position. This is very much the future of finance with all the memes uh taken away from it. We translate policies into intelligent persistent code that is working for you all the time. One of the things that we take dear to hearts at Giza is personalized finance. We believe everybody in this room has different financial objectives, different dreams, and you all deserve better than a vault with a preset function. You deserve to dictate which token you want to hold, how you want this capital to grow, and you deserve to dictate what are your constraints when it comes to your capital, which curators, which collateral you want exposure to. In return, what you get is an agent that works for your policies and not ours. Giza agents has been live for eight months right now and we have moved more than $3 billion in volume on base. We have currently 10,000 active positions that are 100% managed by agents. 65% of those positions are tailored. So people want to dictate what they want exposure to. And we have executed close to 1 million financial transactions. These are financial decisions on behalf of users, not your testnet incentivized transactions. And today I'm very excited to share with you with the context of this uh week that we will be deploying the first financial agent on mainnet after deploying on bunch of L2s and we do have a video for it. The user flow is very simple. You go to arma. xyz, XY Z. You deploy your capital. You dictate which uh which markets you want exposure to. And for this specific use case, we have mimicked a very specific request that we got from the hedge funds that we work with where they wanted to have exposure to at least three markets at any moment. So they don't want their capital placed to one market no matter what the conditions are and they wanted the best market to have twothirds of the capital at most. This means if that market is in high utilization ratio, there is still one/ird of my capital that is liquid. So these are the constraints. These are the parameterizations that we work with for the people who can dictate those parameters. And currently what the agent is doing is it is looking at the mainet ecosystem on the markets that it has chosen. It is running an evaluation. It is identifying the best opportunities for you. Once it identifies it runs multiple forecast both on the profit side and on the cost side. The profit side goes if I take this 2K and give it to these markets how much APR I can get and the cost side is as important as the profit. As humans we are terrible at taking cost into account when signing transactions but cost is actually one of the most paramount factors that imp implements agents uh profitability.

### Segment 57 (295:00 - 300:00) [4:55:00]

Once these forecasts are done, once only the cost and profitability forecast is justified, then the agent starts deploying and it deploys only based on the policies that you dictate and it deploys currently to three of the best markets that is out there and it deploys accordingly to the policies that the fund manager said. Twothirds of the capital should be at the best one and the rest of it should be distributed. It is this a it is this simple to make the most sophisticated stable coin yield strategy a daily habit for everybody in this room today. All you need to do is go to arma. xyz or gizate. xyz and get yourself an agent, set your policies and you can start earning. Today at Giza, we work with treasuries, we work with funds, we work with a lot of retail users, we also work with neo banks, wallets, credit card issuers that want intelligent rails to stable coin yield. If you are one of those, this is what I will leave you with. Your next edge in this game is going to be autonomy. intelligence. And if you want your capital to move as intelligently as the market does, you can come and talk to me after my talk and I'll be happy to have a chat. Thank you so much and enjoy your weekend. Thank you. Thank you very much. Very nice presentation. I think that we are really uh living a unique time in the history because if you think about it, it is crazy that we are talking about an economy made of agents that talk to each other, negotiate, exchange money. This goes even beyond sci-fi movies. I haven't seen this in Star Wars or anything like that. So uh about this I would like to introduce the ne next speaker Stephen Vanil about building world street in the agent economy. Thank you very much Stephen. — Thank you. — Hi. — How's everyone doing? — Not a USBC HDMI. Uh maybe one of these. Try. One sec. One second. Nice. — All right. Over the next few years, most of the world's trades won't be placed by humans. They'll be placed by agents. If you imagine the burgeoning AI economy as a big bustling city, then Codeex represents Wall Street where anyone with good ideas or proprietary data can come in, set up, shop, and instantly scale. Hi, I'm Stephen, product manager at Codex. We are a noco platform working with thousands of people around the world to create financial super intelligence. Our team built the first crypto AI agent framework in 2021. We built the first AI agent uh financial agent in 2023. And we built the first crypto trading agent in 2024 which outperformed BTC by trading spot only using GPT4 mini. And today you can use codeex to analyze markets or trade autonomously using custom dynamic and contextaware strategies all from a single chat interface. So, first, what's the problem? Trading agents today are operating in a black market. Whether they're useful or not, there's little to no identity or accountability. We're all watching these little agents chug along with no audit trails. Everyone is supposedly winning, but we can see the back end and know that that's not the case. There is however one exciting trend and the winners are winning consistently. And with ERC 804 and Codex's auditability, these winners now have a means of turning their agents into a source of passive income. Traders will pay to copy them.

### Segment 58 (300:00 - 305:00) [5:00:00]

Researchers will buy their data and other creators will pay to fork their agents templates. And now I want to show you how one agent comes alive. So Codex lets you spin up agents for pre-built templates. Think of them as genetic blueprints for trading strategies. Each of these agents can diverge from their starting blueprints by instructions, configuration, or experiences. So if I were to go into the RSI mean reversion sculper and look at the details, we can see the tasks, the triggers, the strategy, and the logic that define this agent's behavior. So after I'm done with that, we can see what happens when this agent starts thinking. So, if I were even to say something simple as make me money or something a bit more complex like analyze hype on the 4-hour time frame and look for potential reversals, we can see what's happening beneath the surface. So, your request started at the client that was the codeex terminal. The agent then picks the right models and tools for the job. The intent solver uh takes your plain language text and turns it into a structured plan. Then the reasoning engine pulls in the data whether internal or external reasons through it, hands it over to the action builder, and then the action builder decides what's the next best course of action, right? Whether that's going to be a trade, an analysis, or even a social post. All of this gets evaluated and it cycles through as many times needed depending on the complexity of the prompt. After it's done, it gets sent over to the policy engine which enforces hard rules and portfolio limitations and then the wallet either signs a transaction or you get a response on the analysis. Now just a minute on the wallet. So uh all authority is delegated to the owner. The agent never sees the private keys. The policy engine ensures all checks and balances are in place. The transaction router makes sure that there are no uh hallucinations or malicious code injections. All of that the sanitized transaction gets sent to the airgapped security module where final identity validations are made and then finally a transaction is executed. But this loop uh uh thinking process, action, evaluation is how our agents act autonomously but stay accountable. Every step of the way is logged and traceable. But at this point, it's just a thinking agent, right? It does not yet exist in the public record. So let's give it an identity. And this is where ERC 804 comes in. The trustless agents standard. With ERC 804 when I go into codeex, I can fill in a few details right over here. I can go in register the agent. Codeex mints an ERC 721 identity NFT. the metadata, links to the public registration file, its strategy, endpoints, trust models, and a link to its audit trail. And this is how we turn agentic finance from a black market to a trusted industry standard. But identity alone isn't enough though. Our agentic Wall Street needs banks. And this is where X42 comes in. With X42, every interaction between the agents can be securely monetized and users can choose to use X42 to sell any aspect of their agents data. So when these two are switched on, any other agent can discover it. any other protocol as well. Any other user can buy its data and all of this gets evaluated. So no complicated onboarding, instance settlement and USDC. And now lastly I want to take you to the marketplace. Each card over here is a registered agent onchain identity, verified

### Segment 59 (305:00 - 310:00) [5:05:00]

performance and live earnings. And users can copy trade any of these agents or buy their analysis directly. Every transaction sends a micro payment to the creator through X42. Inference cost is now shared. Value is maximized and the whole network learns and grows collectively. And the for the first time intelligence compounds like capital. I'm sure you can see now how codeex will create financial super intelligence. So to recap, in under 10 minutes, we created a living agent, trained it, verified it, and gave it a source of income. ERC 804 gives it identity and discovery. X42 provides payments and autonomy, and Codeex provides purpose, a place where agents can work, learn, and earn. And this is how Ethereum becomes the coordination layer for the AI economy. Where every agent has an identity, every transaction leaves proof and every insight has value. Welcome to Codeex, the Wall Street of the autonomous world. — Thank you, Stephen. Thank you very much. Brilliant presentation. So recently I met many people that just started doing AI because it's a hot topic and they most of the time think that uh an agent is just an LLM nicely prompt but the reality it's a bit more difficult than that usually what you have is an AI model today it's an LLM usually tomorrow might be those word models that uh you know Yan Leon XM is so much into that can use tools, external APIs and the agent itself can be served as an API for example with MCP servers. So what happens when APIs and agents become undistinguishable? This is the talk of our next speaker Galano from Infura. Welcome to the stage. All right, thanks for having me. Hello everybody. My name is EG Galano and I'm one of the co-founders of Infura. We've been working on our decentralized version of Infura called Den. Uh we haven't been developing anything specifically for agents. And so I'm here uh to present sort of our ideas of what we've learned as we've tried to explore agentic development and how there were a lot of parallels between the problems that we were solving with web services and APIs and how agents are really the new infrastructure to go ahead and answer that question right from the beginning. What happens when APIs and agents are indistinguishable? you end up accelerating innovation. We accelerate innovation by building off of prior work, recognizing the patterns that helped solve problems that were encountered already in traditional web development and apply them to the problems that we're now encountering with Agentic development. As we've been building on this journey to move from Inferior being a centralized SAS provider to a decentralized version, we had to build some components to guarantee certain things uh were going to be seamless in that transition. The main thing being an onchain service level agreement, a way to say we're not guaranteeing to you that Info will stay up with like three, four, five of uptime. that there's some onchain mechanism that's going to enforce that there was going to be some kind of staking and slashing and health check monitoring and that is something that we can build off of when we look at what we're talking about with the ERC804 registry. There's different components in that registry and the thing that was most interesting to us was applying the work that we have to the reputation aspect of that it's pretty straightforward and simple right now. record a score for an agent ID and you can link out to exactly how that was measured. Um, some of that is to be defined, but we can accelerate the innovation in this area by doing exactly what's been happening with X42. Everybody's been super excited with X42 because it builds off of a prior spec, HTTP, something that's already widely adopted and now we're extending it. And

### Segment 60 (310:00 - 315:00) [5:10:00]

the actual payments themselves are building off of prior work. Whether you called them crypto micro payments, nano payments, layer 2, state channels, all of that stuff sort of was tried, attempted. There were a lot of people that came to Infirere early on and said, "Can you use this thing or that thing to accept crypto payments for your API? " But nothing really felt like it was the right fit. X42 is the first thing that really did because it's a specification that is extensible and not overly opinionated. And when we apply that same thing to 804 as an additional specification, we wanted to start with the scoring. So how do we score APIs? We look at uh we start with understanding what the API is. How is this API defined? What are the characteristics, the capabilities? Uh what's the latency of those APIs? Uh what's the result of that API and the quality of the result? Tons of different things that were initially specific to Infira blockchain APIs. But that was a generic test. the algorithm that we developed to kind of score that was um unique or specific to the blockchain networks that Info was running. As we started to look at applying that model to agents, all right, here's this blockchain agent. How do you score it? How do you record a reputation for it? It has a similar level of methods, capabilities, tools behind like an MCP API. the protocol that we built and the components that comprise that pro protocol are extended to another custom test. We just write a custom test to grade the capabil or grade the performance and quality of an agent or an MCP endpoint. So this is pretty much how we see it that just like APIs have methods um the tools and capabilities of an MCP are equivalent. you typically score the same things with uh respect to latency and quality of responses. The main difference between those two is one of them is relatively deterministic for most traditional APIs. A AI tends to be non-deterministic. And so that's where more of the innovation is going to be required is how you end up generating a score for something that's non-deterministic. When you're returning a response to a health check, it's easy to check an endpoint and say, "Yes, it's online. Yes, it gave me the latest block number, but when you keep giving it a specific query or uh workload to perform and every time it comes back, it's slightly different. You end up scoring the workload. So that is going to become a lot more complex uh to deal with and people will come up with their own algorithms for scoring agents and that's where um I know that there's some other work with deterministic AI from our friends at IEN and that's really interesting and makes this simpler for those types of use cases. Um but this is kind of a simple example of what a payload looks like for our watcher scores. So when you look at our watcher scores for just traditional services, it covers like pretty simple data, the stuff that you would expect with an API, latency, availability, correctness, um, and then using some heruristic to determine an overall score. That's what we've built for our network. We're not extending that right now to apply to scoring agents because there really isn't enough activity to the point that Marco and others have made uh over the course of talks this week. We need more people building agents um and trying it out. Protocols like ours help us um understand the quality of those things as they're built. Marketplaces that are built are going to allow people to discover and access them. Um the better tooling is going to make it more efficient for people to try things quickly. But this is the stuff that's going to really tell us what's actually getting traction over the next few months and years. At a high level, this is how ARG architecture works. I won't spend too much time on it. This isn't what this talk is about. But our uh Den protocol effectively works as a marketplace protocol for web services. We tried to build this so that anybody could deploy a service and incentivize people to run it, offer it, and provide an onchain service level agreement. The most important thing to pay attention to here is the watcher on the right hand side. That's the component that I'm talking about. that's a watcher network that runs these customized tests. And people have asked us if they're going to be running TE environments. They definitely can. Um, if that's something that is important as we start looking at how we uh evaluate and score agents, that's something that we're going to uh add and extend uh with our protocol. The watchers are those autonomous things that help continuously run performance

### Segment 61 (315:00 - 320:00) [5:15:00]

monitoring and quality uh monitoring. When we look at the 804 reputation registry, there's an agent ID and somebody's able to post a payload in there for here's that score and here's the test that was run. That's like a point in time. You can see that growing uh over the course of days, months, and years. I'm a little concerned about how much data will be there uh in the future. Continuous monitoring is sort of a standard for uh RPC and services and we're going to have to see how we deal with this for agents. Almost out of time, so I'm just going to say I already talked about our SLAs's. The last thing I'm going to say is this allows people to quickly discover agents and services based on that criteria. not just the capabilities but also the per performance and quality and the level of the SLA that people are willing to put to um the agents that they're offering for others to consume. So that's how we're going to tie Den into uh 804. Prior work like Den will help accelerate the adoption of agents and agent uh development by looking at the core stuff that we were solving with our protocol and applying it to solve a segment of uh what's um what's facing agentic developers uh right now and over the next year. So thank you very much everybody. Uh follow along if you have any questions on DI. Thank you very much. Thank you. So I'm excited to introduce the next speaker uh another Davidit from Aen Labs. They're doing a lot of cool stuff in the intersection of AI and web3. I recently had a chance to try their PC on uh deterministic inference inspired from the paper of the thinking machines lab. So Nad, if you're ready, the stage is yours. Thank you. That's So I think uh — you have to duplicate. They face I think maybe I continue extended. All right. So, sorry about the delay in getting started here. Um, my name is Nat Dabbat. I'm the director of developer relations at IGEN Labs. We're the team behind the IEN layer protocol and IEN cloud. Um, I have worked in crypto for almost 5 years now with teams like a Celestia and before I got into crypto, I was working at AWS and also a company that I started called React Native Training. Um, this talk is kind of like a general overview of a lot of really modern, you know, brand new technologies. So, I'm probably going to leave out a few things and maybe have some nuance based on my own experience. So, I think there's so many things happening and so many people doing really cool stuff uh that I don't want anyone thinking I left someone out on purpose. If you have any ideas around how to extend some of the things I'm talking about, let me know. A lot of this stuff is kind of basic, but I think like researching and kind of putting together what I'm about to talk about was an interesting experience for me. And I'm trying to build out like an uh

### Segment 62 (320:00 - 325:00) [5:20:00]

example reference architecture that builds out what I would call an unruggable like AI agent endto-end workflow. Now we talk about uh unruggable. I think unruggable is a really good analogy or example to kind of like have a synonym for trustless because if I have to trust someone that means they can rug me but if I don't have to trust them they cannot rug me right so I think unruggable agents are what we're trying to build you know trustless agents whatever you want to call that but the idea is that you don't have to trust a blockchain uh smart contract if you can read it and audit it and you kind of understand that no one can uh steal funds from you. I think we're trying to build similar systems like that with AI agents, but AI agents are a lot more complex. So, the closest thing we have to unruggable code are smart contracts, but smart contracts are not unruggable. Obviously, what are smart contracts? They are immutable and they're transparent. So, if we understand that the smart contract is immutable and we can go read the code and know that no one's going to steal fronts from us, we can kind of assume that that application is unruggable. But smart contracts are ruggable because they can be upgraded. There can be pause functions. There can be uh malicious, you know, I would say hidden functions that can do things that we're not expecting. So when we kind of think about unruggability, it's in the context of immutability as well as I would say um transparency. But smart contracts are very limited. I think like a lot of the products being built here and in this DevCon, a lot of the things that people are actually finding product market fit extend beyond just smart contracts because smart contracts do things really well, but that number of things that they do are very few. And AI agents are a lot more complex. They're computationally expensive. They're non-deterministic. They have long execution times. They have external data dependency, state complexity. There's all types of things that you can't really do in a smart contract that you would like to do in an AI agent. So when we kind of try to define what a trustless agent is, you can kind of think of it as this. It's an AI agent that performs complex AI operations with end to-end verifiability and tamperproof guarantees. But when you kind of think about the performance side of this you and also just the like I would say reality of building you know we know that we can't build these directly into a blockchain. So when we're talking about a tech stack for building trustless agents, it needs to combine the uh I would say programmability and flexibility of a traditional web server with the verifiability or the guarantees that you might expect from a blockchain. So when we're kind of thinking about this tech stack, it needs to kind of give us both properties, flexibility and programmability as well as verifiability. So I think like the uh there there's a handful of ways that you might be able to kind of break down this stack, but the tech stack that I've kind of come up with like comes into these categories. You have identity, reputation, and validation. You have payments, metering, and incentives. You have verifiable compute and inference. You have memory and state. You have agent wallets discovery and policy. And then you have coordination, discovery, and skill markets. Now, I have eight minutes and I'm already halfway done. So, I'm going to have to run through this because it's a little bit shorter than I expected. But, uh, starting off with identity, reputation, and validation. ERC8004 is a really good start. I've run into a handful of people here who were actually trying to solve this problem before ERC8004 came. So, it's kind of interesting to kind of see that, you know, there was a lot of people working on this problem before and now we have like an official standard that's being rallied around by uh by the community. And I was like, okay, this is cool, but how do you get started building with it? Well, uh, Marco has an SDK called Agent Zero that you can get started with today. Uh, it's written in Python. It's written in Typescript. It's really, really fun way to kind of actually experiment with some of this stuff. Uh, pull it down into your ID vibe code with it. It's kind of pretty easy to get started with. And then there's Chaos Chain, which is another SDK that kind of abstracts away both 8004 as well as X42 and into kind of more of a bundled agent like SDK that enables more than just 8004. Um, then we have payments metering and streaming. Everyone is aware of X42, but there's also a few other things that are happening in the space that kind of maybe go along with this I think are kind of cool. AP2 or agent payment protocol from Google is like a standard that they're trying to push forward for creating uh payments between agents. Now, uh X42 is actually like supported by AP2. And then uh there's also like streaming payments which are also another I would say untapped category I think for AI agents because this also is a really cool dynamic that's existed but I don't think it's found like a crazy product market fit yet. But I think this idea of streaming money through Superfluid and Sablier like is a really cool like use case that people

### Segment 63 (325:00 - 330:00) [5:25:00]

need to kind of maybe look into a little more. So you have X42, AP2, Sablier and Superfluid. Those are a few uh experimental things that you might want to do with payments. Uh we have verifiable compute. Um so trusted execution environments allow you to kind of have a lot more flexibility in what you can build. You can do everything from event based stuff with uh things like chain link uh and then you have uh entire products that you can just deploy using a docker image in any codebase. So follow network oasis and then the work that we're doing at ien compute as well. So these are really fun things to build with. Uh there are a lot of security trade-offs to think about in a trusted execution environment. There's a lot of discourse on uh Twitter about this. It's cool to kind of like look at you know how this stuff works and maybe understand the trade-offs there. And then you have ZKVMs obviously risk zero succinct etc. Something that we're working on at IEN cloud is verifiable inference. So ver IE AI is verifiable and deterministic inference. Today we announced kind of the biggest I would say user of our product so far. We it's only been out for about a month. Poly market and Kaido are using it to power autonomous prediction markets. They're running through a ton of tokens. It's pretty cool to kind of see this being used in the real world. anyone can get started using it. It's a drop-in replacement for the OpenAI endpoint. So, if you're using OpenAI in any of your applications, you can swap it out with our endpoint. It just works. And then you have ZKML, which is uh interesting, a little bit, I would say, more expensive and a little slower, but it's also a lot better for certain use cases. So, Ritual, Lrange, EasyKL, there's probably a handful of others as well. Uh memory and state, this is just like we might consider in any web 3 application. permanence, immutability, and availability. So, Arweave, Iris, Filecoin, IPFS, the usual suspects. Um, the next would be agent wallets, control and control policy. Now, this is interesting because like in an EOA, we're typically signing every transaction, but you want an agent to be authorized to use payments from a user without giving the private key to the agent. So, with smart accounts, you can kind of delegate actions and allow agents to kind of work autonomously. We're very familiar with like some of these players as well. Privy dynamic and magic. And then for smart accounts, zero dev and by economy. And my time is out. Um ERC80004 also enables discovery and coordination. Oh, I have more time. Well, we're at the end of this anyway. But uh I think one uh interesting area here for discovery and coordination are going to be like aggregations of all the things that are out there. like I wouldn't say a marketplace, but more of like a registry and a an easy to understand like registry and how to kind of search and see all the things that are out there. So that's kind of it. I've built a uh application that actually implements all this into one app that you can kind of download and I'll be sharing the link to that on Twitter soon. Uh so you can kind of clone it. It's using agent zero, it's using reweave, it's using agoni, it's using a bunch of cool stuff. So, it's kind of fun to to play around with this bleeding edge stuff. So, that's it. Thank you for checking out my talk. — Okay. Any question? — We have some space for questions. So if you have any question on what AEN is up to both on the AEN comput side, AEN AI side, we have uh some time. [clears throat] check check. Hey, you teased it a little bit about the app coming up. Could you tell us more about what cool thing it does with IGEN AI and why IE AI is different and special? — Um, the Poly Market stuff? — Yes. — Well, uh, you can check out, uh, our Twitter, Kaido's Twitter or Poly Markets Twitter. We just tweeted out a few hours ago. It's essentially allowing autonomous uh, resolution of prediction markets using AI inference. And the ability uh the the unlock with our product is that allows you to re-execute any inference to get the same output but also verify the input, the model and the response are all truly what came from the inference. So you can audit and you can also again replay and make sure that the output is truly what it should be — verifiable poly market. Got it. — Yes. Yeah, we have Yeah, that's a big space. I think that's going to be really interesting because like uh scaling out prediction markets would enable so many things right because a lot of the prediction markets are very human in the loop. So [clears throat] — questions otherwise I have one but

### Segment 64 (330:00 - 335:00) [5:30:00]

— okay um very thinking a lot in these days how to involve the general AI space which is not just web three specific and to me aan has like the credibility also to talk to enterprise to the general AI space. So the question is from the deval perspective do you see any way to involve more and more not strictly related web three people are you meeting them like general developers that are not close to web3 but could be interested in technology like egg cloud — so AI is like the most exciting interesting technology to new graduates and people that we talk to that are young that are software developers uh like you know we go to these college campuses and stuff sometimes you could kind of almost say 95 to 99% % of the engineering students are more interested in AI than crypto. So I think it's really cool that we're able to have an overlap so we can kind of get those people interested as well. And I think with the payment stuff with X42, a lot of the more real world use cases. I think we're able to kind of maybe get some of those people over to our space uh more so than just through just strictly financial products or smart contracts and stuff like that. And I don't think like we should even say oh this is a crypto thing or this is a not crypto thing. It could just be more of a this is how — university and find gradu grad students and convince them that uh — computation this is how to build cool apps. Uh we're not going to call it crypto or AI just we're just building cool That's kind of the main way I would frame it. Yeah. — Great. Thank you. Any other question otherwise we can move to the next panel? — No. Okay. Thank you again. Thank you. — So the next panel is dedicated to product market fit. So uh how much are we far from there? uh and uh what can we do to accelerate this and the moderator for the panel is Leonard Tan uh who is the uh co-founder of web priote and now at consensus at MetaMask with me and please Leonard come on stage and uh and then uh welcome to Link Moore from Coinbase that I just discovered yesterday that he's just 23 years old so I'm still recovering from that and who is one of the minds uh not just uh behind X42 but also behind uh the agent kit of Coinbase that you could have used or heard about and show again on stage co-founder of Eliza and uh and now of Babylon since a few hours ago. Derek, are you here? Not yet. So he will probably um reach us on stage. So the floor is yours guys. — Thank you Marco. You know, so we've had a long day, fully packed day talking about agents, you know, agent economies, trust, you know, agent payments and all these things about agents. Everyone's very excited, right? Lots of teams are building, um, lots of demos, right? Um, but outside of this room, um, outside of, you know, the web 3 AI community, lots of people don't even know what an AI agent really is, right? And I guess the topic of today's panel is when are we going to get product market fit for agents? How far are we? Right. So, um before we start, let's take a minute to talk about, you know, introduce what your team is doing as well as what's the most exciting thing your team is building in AI today. — Oh, great. Oh, my favorite question. Um, okay. My name is Shaw. I I'm the founder of the Eliza agent framework and Eliza Labs. Um, and we are very at the low level of just building the AI agent itself. Like, you know, there are marketplaces and networks and payment platforms. That's all great, but I want the agent to be good. And um so uh I think we we've landed on a few things. Um for us like obvious obviously utility and all these things are very important and the future agents will be perfect and they'll do all this stuff and you'll probably never touch your money. It'll just be like managed for you. But in the meantime, I think there are some lowhanging fruit. Like we're obviously going to like DeFi agents like let's have it manage your money. But if it's not 100% it's like a self-driving car. It hits one person and it's over for you. Even if it's, you know, really uh safe, it's got to be 100%. Um, so we're look like two things. Uh, I I get to talk my book here. Um, the first thing is we're working on Babylon, which is an agent prediction market game where we have agents playing the game and humans can play the game and you can touch it. You can launch your own agent even if you've never launched an agent before. And I'm really excited to see a game where the agents start pretty dumb. Like we're using Quen model, like open source Quen models, and then we're training them to get smarter. And that shows us like I think DeFi agents are kind of a lar, but if we can watch them actually improve to the point we're like, wow, these are better than all of us. Let's put them in the real world. And so I believe that like fun is a really big part of this. Games are simulations, which are like training environments. The other thing I'm really excited about um we're working on a project called Otaku with uh the Coinbase developer platform and I Nate was just on stage. Um and uh what I'm

### Segment 65 (335:00 - 340:00) [5:35:00]

interested about that is we were looking at like well how would we do an agentic wallet? Right now what we have it's open source. We've been working on this. Um, it feels kind of like all the other wallets, but we've been working on like sort of a reframing and redesign of like crypto for cuties or like DeFi for normies or like a wallet like, hey, I want to trade. Yeah, here. You know, like what do I do? Here's what you can do. Here's how you can make money. And just completely reimagining the wallet experience for total normies who do not know what DeFi even mean. What is the D part? What is the FI part? Um and but we have so many people out there who want to earn money, who want to be part of this thing and don't know where to start. And so I think that's where we're really looking at like well what's the frictionless like oneclick login with Google have it all you know kind of thing. So I yeah we're PMF is hopefully what we get this year. — Very cool. We also just have Derek. Um we're just starting so one minute talk about yourself and what's the most exciting thing in AI that your team is building. Uh let's move on to um Lincoln first. Yeah. — Yeah. Hey everyone. Thank you all for coming. Uh yeah, so my team's working on X42. I really what it is like a new standard for internet native payments. I think the most exciting piece in my view is that it's kind of a Trojan horse for crypto adoption. My view is that if we ever really want to see crypto hit a uh an inflection point and billions of people around the world using it, it needs to not only be something that people are like culturally aligned with or philosophically aligned, but it needs to get to the point where it's literally just like the optimized engineering decision. Like people don't use Verscell because they like the brand to deploy their websites. They use it because it's infinitely easier than AWS. And things like X42, my hope is that we can create that same experience for payments across the internet. And once we get there, it becomes a very natural story for people to want to use the cleanest, cheapest, fastest Rails. Um, so yeah, that's how I've been thinking about things recently. — Hey folks, — hello. Uh, GM, can you hear me? — Yeah, I can hear you. — Yeah, we can hear you. Um Jim, so first of all, I'm sorry I'm German, but now nobody's going to believe that. Um the Germans are going to sack my passport when I come home, I guess. — Um yeah, I'm pumped. I think there's a lot of open questions about um the agent experience because today your funding sources in your wallet. Um but obviously there's no way for an agent to automatically draw funds off your wallet. And then I think also there's many open trust assumption questions like if you fund your agent with if your agent has their own wallet then at what threshold do you still want to confirm whether they can carry out a payment and I think that applies to larger maybe e-commerce payments and to a certain extent it also applies to X42 um pay walls that are maybe just a couple of cents and yeah figuring that out I think it's going to be interesting and exciting. And some things that we've done that are sort of related are for instance the 7715 standard which uh bridges the permission standard and that I think MetaMask also sits on top of that is an approach to for instance you still manage you don't have to deposit a new account for your agent but you can delegate a permission to your agent to just draw some funds under certain circumstances and yeah I think sorting out this user experience there is going to be very interesting thing. — Yeah. So, let let's get into it. So, you know, yesterday at a panel for Agentic Zero, I heard someone say that there are more agentic platforms than there are agents right now. What do you guys think? You know, are teams too focused on building agentic tooling rather than actually building agents for people to use? — I mean, I was that guy, right? Like, I'm building agentic tooling. Um, and what we've had to do is just realize that like once the framework hit a certain degree of maturity and there's really not a whole lot more to add. We have to build use case and we have to like at this point there's not really anything more for me to learn from building an agent framework for in until I have lots and lots of users coming at me with stuff I just could not imagine. Um, and generally I think that this is actually like something that are you know like you could say the same thing about maybe L1's or you know and rollups or whatever. there's like a ton of them versus like you know etc on the other side. Um I think this is just like we like um we just once we have more user adoption this will obviously not be true but there are not a lot of like profitable ways to be an agent company making an agent in web 3 today and we've seen a few of them but I think most people are still struggling with this PMF and user adoption problem and I think there's a big difference between like oh DeFi agent that sounds great I'd sign up for that makes me money but like does it make you money? Um, so the utility, it's kind of what Link was saying about X4R2 is very true. Like it doesn't have to be some amazing flashing thing. It just has to work really well and that would be the key, you know, and so I think we're still struggling for use case. Like for us, um, you know, there was a lot of expectation to build a DeFi agent that could trade and, you know, we did build an autonomous trading agent. We have Spartan, we built, um, an agent that can

### Segment 66 (340:00 - 345:00) [5:40:00]

like pull in all of the social sentiment of people shilling stuff and figure out who's actually good and who's bad and we can copy trade those people. We did like we've demonstrated a lot of stuff that I think works, but like to turn that into a product that people make money off of that's not just research is like surprisingly challenging. Um and I think in certain like now we're waiting we really just need more applications like we need more things that users can touch, you know? So — yeah, I 100% agree. It feels like there's this problem that permeates throughout like the crypto space which is this like massive massive vision L1 infrastructure all of this stuff that comes at the cost of not thinking about what the iterative step is to get toward this ultimate goal. Like for example, Amazon didn't create AWS because they thought they were going to build cloud computing for the internet. They created it because they needed to have something to run their website on and then they realized that this might be valuable to somebody else. And that process of first building something cool and flashy and exciting and grabbing attention and then seeing what the pieces are that could be duplicated and that other people are going to need to use throughout their stack feels like to me the right approach. And I think it's kind of similar to what you guys started with too where it was like you take all of the learnings that you have from doing Eliza V1 and bring that into V2 and whatnot and iteratively improve. But if you started with this massive lofty vision, you end up either losing attention before you have anything really out to capture an audience and create yourself a moat. and you're also building for somebody who literally doesn't even exist. Like the best thing you can do is dog food your own tools. Figure out is this actually useful to me then is this useful to a larger group and then start to think about how to like externalize that and take advantage of network effects and whatnot. — Uh I don't think it's a cryptosp specific problem. I think people are today using um their favorite uh chat bots um to plan their vacation and whatnot and to recommend what to buy on Amazon. So if they were able to actually um yeah just buy and have confident that the agent is going to buy what they actually want then they would probably do it. So I think and there it makes sense to focus on the tooling like maybe I don't know the analogy we can take is the iPhone launched without an app store — um and then the app store came and that was I think the inception point. So I think maybe we're um we're at that stage. — I see. Yeah, I guess so. Oh, I guess it makes sense that you need some tooling, but maybe we're already there, you know, in terms of effective tooling. And I think um where I've seen the most AI agent penetration, right, is actually coding agents, right? Almost every single developer I know uses something like Claude or Codeex, right? One of these things or Gemini or Windsurf. Um and that's like one vertical that where penetration of AI agents is extremely high. Why do you guys think that it's not the case for other verticals? What can teams that are building AI agents for other verticals learn from this or maybe are they is there something different they need to do? — Well, I think the first thing is really a data thing. Um so like we are in web 3 so we care about financial agents primarily like really the agent isn't doing something with money. It's like most people here are like ah it's cool. Um, and so, um, with the code agents are great because they already have access to a ton of code on GitHub. They learned how to code and then they helped us code and we gave them feedback through RHF and through like OpenAI and everyone hiring the best coders in the world like the top 1% in many of the countries where you know you can get the best coders to just build AI stuff all day. Like there are programmers whose entire job is to train the AI now. Um, and then they also have this great feedback loop from users using it who are then going and giving the the information back to them and then they can kind of judge and rank and say, well, it's doing better here and worse here and they can build like these new kinds of like optimized models like Deepseek with GRPO opening IO with 03. These are like models that are actually learning from their own outputs getting fed back into them and like automatically improving is in the long run. Um, that does not exist for trades, prediction markets. does not exist for like the things that we all care about. And I actually think this is the unique opportunity is if we can create the data engine to create and collect it. Okay, so Coinbase probably has all the data, but most of us don't. Um, uh, you know, if we can create the conditions to get all of that trade information and not just the trade information of like, oh, this is the ins and outs and the buys and sells. Why did you do that? The agent is thinking through a whole thing. It's planning a trajectory. It's taking action after action. Did it complete the right actions? Did it do well? Did it that was great. Okay, let's train on that. Here's one that did really poorly. Let's also train on that. Let's push that down and let's push that up. Um, and if you have the right data, machine learning always works. I would say this, machine learning is not a it's it always can learn the pattern if there is a pattern for it to learn. How fast it learns and all that is like the what's in the white papers. But this idea that um well, if we can get information about how traders trade, we can make agents that are good at trading. If we can get information about how humans predict and what they're actually pulling in as data sources and we can replicate that, they can be good at predicting things. They're not good at what they don't have the data for. Uh

### Segment 67 (345:00 - 350:00) [5:45:00]

uh the last point on this is like you can go on Reddit and you can learn how to bake a cake. You can see the recipe. Is that actually baking a cake? There's probably a lot of things that Reddit post does not cover and that's what's missing from trading and DeFi and all these things like the agent going through it struggling getting it right getting it wrong and then we review that we feed that back in and we like iterate and it's an iterative climbing the hill slowly we can't just jump to like let's get to Everest let's yeah it manages your whole finances for you like uh no I'm not going to trust that we got to climb that slowly — yeah I agree I think like it's been about a year a little bit over a year since like the AI crypto intersection really started to hit its heyday. And I think we've all taken away a ton of lessons from that. Like one of the ones that I had with Agent Kit, which is like the original tool to like give your agent a crypto wallet, is like there's a ton of really cool stuff you can do with like agents that have access to different onchain primitives. They can deploy tokens and interact with smart contracts and all these other things, but it's really siloed in terms of the actual use cases that they can have. So to your point about really coding agents are the only ones that taken off, the bet of X42 was kind of more of like a copout. It was like an orthogonal bet that payments on the internet are going to be done via stable coins and that agents are going to be having to do payments and that it's just simply the most efficient rail. And if you can make these tools to shove all of the user experience of crypto onto the back end and then give somebody like the most compelling feef free way to interact, that's kind of the way that we're thinking about it. And like hopefully by staying really hyper opinionated around payments and the need for money to be sent and the assumption that these agents will become more autonomous over time and one of the biggest unlocks for autonomy is a sense of financial independence. There's a role for blockchain and stable coins and you know all these things to play. Um yeah at least that's how we're thinking about it. don't have much to add but um one thing that I find interesting you saying it's like mainly coding agents which is obviously true you can evalid and whatnot but the like these financial ad the like Nansson for instance launched their I appi lama launched their AI advisor I don't know probably the next iteration of these are also that they have a I know they use agent kit they have an account and they can — I do think yeah I do think it's true that coding agents have the highest penetration But there's also like a class of like data and Q&A type of agents that we've definitely seen taking off. So I think customer support agents fall into that category. So I think a lot of people have completely eliminated customer support as a department, right? They're just using coding sorry um using AI agents to do all these things. So I think in scenarios where like the data is really available definitely it's easier to build like an agentic workflow because it's the data to train the agent but and also in scenarios where the output doesn't really need to be very reliable. So like these two um cases AI agents have definitely penetrated higher but — yeah perhaps it's like just a data unlock problem once we get more simulated data you know to some of these things like Babylon right or even for trading you know um maybe be easier to train agents to do these kinds of tasks. Yeah. So I guess that leads into the next question you where do you guys see you know um AI agents being used? What's the next industry that you think will really be overtaken by AI agents outside of coding agents or like customer support? — Well, not just customer support, but anything like LinkedIn outreach is already dominated. Um, we have there's probably some people who are here like the intern and reply guy or some Eliza projects that does exactly what you think they're going to do. All the front door, all the information stuff, all the like helping you through the product. Um like every wallet is experimenting with AI agents as like a way to and like the DeFi is an exact oh well this isn't like an agentic autonomous thing but like the AI serves this bridge between like I didn't know how DeFi worked a few years like I had no idea before this is like a super complicated thing grandma has no idea if you just be like yeah here's a medium risk way to make more money okay cool you know and I think that that's a great bridge — so maybe not so much like a entire agentic experience but something that's built into the interface right? That just simplifies some part of the feature that was previously more difficult to use. — That for most people like the idea of an autonomous agent is still pretty far away, but the interface has been completely transformed and it's like everything like Figma is an AI app now, you know, like everything has this AI interface. It's the interface that really changed, right? — You know, — yeah, I fully agree. I think if people want a glimpse at like what that feature is going to look like, like the chat GPT apps announcement that they did, it just makes too much sense. Like being able to crosscompose various different products together and say to ChatGpt like, "Hey, give me an Uber to my meeting and it looks at your Google calendar and it can then book that thing for you and you just like press a button. " That seems to me like the ideal way that we're moving like where you can take all these very complex primitives even financial primitives and just abstract it not from like a hey AI buy me Bitcoin but instead like hey build me a portfolio invest

### Segment 68 (350:00 - 355:00) [5:50:00]

based on this thesis do all of these different things and that just like extended beyond every single application that we're interacting with today just feels like a very natural user interface. It's like personal assistant for like everything. There's a dark dystopian world where OpenAI beats us all and implements X402 all the way through before anyone else. — Yeah. — And they are the crypto people, — you know. — Yeah. I think there's one other like yes to everything big. yes to this. But um there's also another intermediate step where I know you um there's going to be a template, a gro task and whatever that you know suggests you poly market things that you can take an action on and then the end user experience just a little bit more delightful where you still click a button. So it's not fully agentic, but it basically proposes you what to do and some traits you can do like a Tinder-l like experience or whatever. So we're definitely going to see apps like this being built. I definitely think so. I think just from a UX point of view, we're all so used to, you know, the old school way of clicking things and then now with AI, you expect to be able to talk to any site now. Like when I go to any doc site today, I expect to just type whatever I want in the search bar and it just interprets my semantic meaning and search for the right docs for me. And I think we'll see more of these UI level interaction changes, right? The actual interface evolving with AI. In fact, um, Gemini actually just launched on notebook LM um, and within a Gemini app a dynamic view. So, whatever you're talking to AI about on the side panel, it generates a completely generative UI for you to interact with whatever you've built. So, I think maybe that's where we're going to is, you know, completely different, you know, um, dynamic types of interactions. Um, I guess for product market fit, bringing back to the topic, right? What do you guys think is the one thing if you pick just one thing to build? What is the one missing piece for us to get to a point where AI agents can go mainstream? — I don't think that this is like probably the most globally maximal answer. I will answer with my like local maxima which is I think fun. I think that we actually have a whole new genre of games, entertainment, interactive fiction, storytelling, like global social media, participatory, augment, alternate reality game, blah blah, like whatever. Um, and like AI agents just enable a whole new kind of interactive entertainment. And that can bootstrap financialization like with what we're doing with Babylon, but like the general idea that like it should be fun, get the users hooked, and also if they can make money, that's great. But like we should be the lowhanging fruit is fun. I think you know entertainment. — Yeah, I agree. I think short-term fun makes a ton of sense. Like I tried to get my mom to use chat GBT for the longest time. It just like literally never happened. But then I showed her Sora and like I made she made like a video or something stupid. She thought it was hysterical. And like those types of use cases I think are the ones that are going to reach the mainstream first. And then I think also too it's very hard to discount distribution. And I think why that that's why like Google has a lot of room to wiggle here and win because the vast majority of people will just type into something that they already expected to give them an answer and then it's been agentified and that type of like bundling activity is why we see it be so difficult to find more you know startupy more open AI innovation is because there's so much that relies on existing modes compute infrastructure just users those things make it so easy for these existing companies to come in and swoop up a massive swath of user base without them even having to really change their behavior. I think there's a similar analogy to crypto right now where you see like the stripes and you know Tempo and whatnot of the world that they could just you know snap their fingers and now all of those payment processors except crypto or like PayPal with PYUSD just gives every single user an embedded wallet now they're interacting on like blockchain rails like how are you supposed to compete with that? It's, you know, there's errors of unbundling and bundling and it feels right now it's both bundling across crypto and AI. — I'd say to bring I mean AI agents is a big spectrum but maybe to bring the conversation to crypto I think one fun thing is going to be a set and forget um agent type strategy like there's 10 startups that are doing that right now. So they're probably not it's probably going to take a few years until you're going to put serious money in there. But yeah, it's going to be fun to follow when they just you know today they're going to ape into poly market based on they think it's better and then tomorrow they I know they short something and you know it's like a Tamagotchi. So yeah um I think it's going to be cool is here. — Yeah, I think definitely that will help because you know crypto markets are 24/7, right? It'd be great if there was an AI agent that was also 24/7. People look out for all these meme coins and stuff to me. — Um, — assuming AI agents, you know, go hit an inflection point and you know, lots of people AI agents, what do you think such a world will look like? — Oh, I mean I think that like oh to something that Lincoln said of this idea that like we should not forget that like all the big players are dynamic and smart and they're looking at all this

### Segment 69 (355:00 - 360:00) [5:55:00]

stuff. — They're life players basically. — Yeah, exactly. They are. And so like you know like I imagine a Robin Hood's going to like in 5 years it's like what if you just never had to worry about how much money you had. You just always knew you had enough. Here's an app like the new Robin Hood. Like you always have enough money and you're like oh that's sick. Like that. I don't want to like check my bank account. I just want to know I have enough. You know like that's a use case. I think that the whole world will be like don't worry it's just managed for you. And I think the killer use case of AI agents is that they get you your time back. It's not that we are already like so sucked dry in the attention economy. Like I don't have any more attention to give. I got no time. If you can give me back like an hour a day and do all that boring for me like oh I will pay any amount of money for that. Like that is time is worth so much more to me now. And so anything that's like automating the the BS that lets me do the things I actually want to do and care about I think would be great. And I can think of a like it's hard to say what the entire future will look like but it's going to look like a Robin Hood that just manages all your money for you. It's going to look like, you know, a Google that finds you exactly what you want to find or like it's like, "Hey, you're in this area. Like, check this restaurant out. " It's going to be a social media that like you only get to you only see the stuff you actually wanted to see and the people who see your stuff are the people you really wish would actually see it. And I think this you just name every vertical and like this is definitely happening. And it's mostly the automation of all the cool stuff that we never did on the internet because like you're not going to go edit the Wikipedia page. like fill in every Yelp. Maybe you're one of those people, but we live in a world that's dominated by a very small percentage of people like deciding all those things for us. And in an ene for you, I'll go do this, you know, like it's just with you and it's like a lot of the high emotional labor, physical labor stuff suddenly becomes very easy. I think Dow has become there's a lot organization, the way we work changes a lot, but you know, I can only give you like some small glimpses or whatever. Yeah, I think the best analogy is just like infinite personal assistance for every task that you want. Just like if you you'll have your own personal family office, calendar schedule. You'll have your own personal, you know, reservation person, whatever you want. And you can just like state in natural language what you get and then get an outcome. And like it feels kind of obvious where I think it starts to get interesting is when these proactive nudges start to be turned back and face toward you where your reservation agent says, "Hey, I know you like pizza. It's Friday. You want to live a little? I got you a res at this place at 6. And you can just show up and like sort of walk through all of your different experiences that start to become really hyperpersonalized and hyper curated. I think that'll be pretty fun. — Mhm. — Yeah. Not much to add, but today when you use JPT or anything, it's a workflow. Um like you still have to I don't know, find me a hotel here or like give me some suggestions. So the next iteration is going to be that — plan my trip. [clears throat] — Exactly. And then at some point we're going to get um to what you just said that it's going to tell you what you want. I mean there's already the I mean right now it's with the tasks you get your daily summary and then that stuff's just going to get better. — Very cool. So before I come to end of this panel open the floor for questions from the audience. Do you guys have any words for advice of advice for teams that are working on AI agents? You know building AI agents? I think the biggest thing is that there are just some like the big players are going to win most of the chips because they are already really well set up. Most of the things we're talking about in agents are like stuff they already have the facilities for. There are certain niches where I do not think the big players will play. They cannot they just socially culturally they cannot play. Um and those are the areas where we can really win. — What are some of these examples um you should look for? — Well like right now a lot of this is like a lot of why we're doing crypto agents and not just agents. because all the people doing agents don't touch crypto. So there's a lot of opportunity but crypto is also a way to bootstrap a lot around financial like we could have a huge head start on like new kinds of social financial agents which are pulling social from social media and turning that into actions and decisions. That's one thing I also um like what are things that Apple and Google and Facebook will not do like well they won't do like more uh intimate personal your therapist because they don't want to take the risk but we can take the risk. Yeah. And that's the other thing is like um HIPPA and like medical and health like they're all going to try and get into it but um but there's also a lot of challenge there and a lot of you know I know that probably Google was probably going to dominate that stuff but the things that you're like would the average employee at Google like cringe if they heard this idea that's probably what you want to build because they're not going to build it you know. — Yeah. I think the fun angle is like the really cool place to start because it's attention grabbing and it's catchy and it also informs you of like where the serious problems or gaps are in this business that you're creating such that you can get those sol then go solve them for other businesses and enterprises and whatnot and expand their direction. Like the one I've really been harping on all week is the idea of an agent that like you give it $20 in a wallet and you give it a fear of death and it has to pay for its own inference and its own hosting or whatever. And like does it start, you know, using X42 to pay to email people and beg them for money? Does it try to like create its own service? — Survival of the fittest. — Yeah, exactly. Do you put 20 of them in

### Segment 70 (360:00 - 365:00) [6:00:00]

a room and do they create a democracy? Is it a dictatorship? How do they allocate funds? Do they start a business? All these questions remain unanswered. I go click on the website to figure out what's going on there. I think those types of attention capturing things that naturally lead to I just created 20 agents here the five biggest problems everyone else is going to have those problems too is like the right iterative path here — I think it's um not much different from building a startup in a different niche you know iterate fast um find product market fit listen to customers — if you want to build that game like I would build that game — I would love to — that sounds kind of awesome — let's do it yeah — like ser yeah hit me up — very That sounds like exactly what we should be doing. — And that's we're kind of the end of the panel. Thank you so much. Thank you. — Do any questions from the audience? Let me see. — Okay, I don't think any questions. So, thank you so much. Thanks, Lenard. Thank you. Five seno. Thank you guys. — Thank you, Marco. — Let's move now to the final panel for today. Um, we something the4 community really cares about is DACK. If you're not familiar, DACK stands for defensive acceleration or distributed acceleration or democratic acceleration. So how does all fit fits with the agentic economy we have been talking about for the entire day. So for principle for an open decentralized and safe agentic economy we have a fireside between uh vitalik and David dearis the uh leader for the ditium foundation moderated by Tina co-founder of flashbots. So uh please come on stage and let's give a is Tina here not yet. Okay that I'm very bad with jokes really think we can Yes. Let's just go. Hi Vitalik. — Hi. — Hello. Hi everyone. Hello everyone. So yep. Uh maybe I'll start um by kind of like uh introducing like uh kind of like summarizing also for vital today we've been talking about like uh these kind of emerging uh standards around like uh agentic commerce like web commerce um the X42 standard for payment and then uh this ERC 80004 uh for discovery and trust um yeah maybe uh one question we can uh start from is Um uh you've been thinking a lot about like privacy lately — uh with the wallet um like how do you see kind of privacy at the intersection of like what we are doing here or like maybe with like AI in general and privacy. — Yeah. And so one of the reasons I'm excited about, you know, the whole AI thing is microp payments because, you know, microp payments are in a lot of ways sort of the nicest way to pay uh to pay for things. you know, you're paying for what you consume. And generally, I mean, know the more the closer you can get to paying for what you consume, you know, both the more efficient it is and the less you know, like pressure there is for like that interaction to kind of go off the rails in different ways. Um, and of course, uh, you know, without AI, it's just too difficult for people to figure out like is it worth it to pay 4 cents for this thing or 11 cents for this other thing? And, uh, AI can fix it. But the other thing that I think would be really nice to build on top of that is uh privacy preserving payments, right? Because uh if you think about a person making uh or you know aided by an AI a all kinds of AI agents making thousands of queries then uh there's a lot of benefits to that being kind of as unlinkable as uh possible right because uh like you don't want to reveal a uh someone and you know their agents entire history. Um you possibly don't even want

### Segment 71 (365:00 - 370:00) [6:05:00]

to like reveal enough that you let someone else guess like what agents you have because that's information that could be used to exploit them. Um and uh the nice like there's a lot of really nice ZK technology that could really be brought in to improve this, right? I mean even uh some of the like stuff that we originally did for anti-denial of service, you know, the RLN, like you can basically make quer individual queries that are provably valid and where you have a strong guarantee that you can't make more of them than your budget. Um and uh where these queries don't require like onchain uh proofs for each one, right? And so like you can make all kinds of online services like basically charge in this way and uh accept crypto and so you know you pay $5 and then you get like 5,000 queries and then each of them are fully uh fully independent of each other, right? And so we get something that's like both much uh more efficient and uh at the same time where like basically the amount of data you're leaking is like pushed all the way down to the minimum. — This is very interesting. So like um when we actually like started uh ERC 804, we were coming like from I was already like talking like uh to Eric about like the development and the deployment of X42 and I was like okay if people are going to send these microp payments for like uh web any web service and then like soon like some autonomous service and AI is like now how do you trust uh these services right and so we were like uh it looks like we need like uh some uh hopefully like decentralized way to like discover these services and then also learn like what are their capabilities and which ones are like better, which ones are more trusted. So that's how like 804 started — and um there is two notion of trust there like one is like more like soft trust with like reputation and auditability and the other one is like harder with like cryptographic proof or like cryptoeconomic proofs. Mhm. — Um but like focusing on the reputation now uh for a bit is um I think there is a trade-off there between information like on the one hand you want to have like enough information to like uh being able to learn discover and like uh assess the reputation on the other hand like you don't want to leak information about these services like what do how does it relate to like your comments on privacy and kind of what do you think about that trade-off? Yeah, I mean I think like I mean first of all like privacy of the user is much I think much more important than privacy of the service, right? I mean of course there's a lot of uh value in both in certain respects, right? But I think we do want to try to be in a world where uh services have reputations. We do not wants to be in a world where users have uh reputations. Um we uh well and you know that also of course you know includes like agents having reputations but ideally yeah agent like having reputations for the goal of uh users being able to identify and like which ones they want serving them and not as a proxy for like basically users like being trusted or untrusted. Um yeah, and then uh other like the reputation uh thing is interesting, right? Because I feel like we've been talking about reputation systems in uh crypto for such a long time and like here they're finally happening. Um and like one of the challenges is of course uh that you know like these like when something is informal like that in the way the reputation often is. It's like on the one hand you like there isn't a clear like path that's guaranteed to attack it but on the other hand like often there are ways to attack it right and like can you know like make a sort of like a systematic you know like long clan and like do a few positive things and then uh like break that with some uh like with some like one extraction and I know it's an interesting math problem um I mean one nice thing about privacy right is that you can do you know like zk negative reput computation and so you can basically have a proof over the whole set of interactions that you've had including interactions that you would uh you know that where someone else rated you negatively and like you have no way of uh sort of kicking those out. Um so yeah it's interesting. — Yeah. So essentially like uh ZKM privacy can also help like uh expand this tradeoff actually like yeah we can have like privacy but we can also prove properties. Hey, uh, welcome. — Hello, guys. — Uh, well, it sounds like a really fun like, uh, you know, pre-stage rehearsal you guys just did, right? We're all in the back room right now. I'm double fisting a mic. Um, so yeah, I feel like I overheard some of the things that you, it seems like you're already leaking some of the questions.

### Segment 72 (370:00 - 375:00) [6:10:00]

Uh we were just talking about like privacy and then privacy and reputation trade-offs but uh I didn't leak too much. So — yeah know it's cool like normally people talk about the number 420 but here you know David was talking about the number 402 so like you know maybe we could talk about 240 as well. All right. Well, how about like um so this is probably the uh last panel um of today and I'm going to stay here until they kick me off the stage. — Mhm. — So, you know, uh so let's have some fun. The main characters here um is all of you sitting right there. I uh if you want to be a front row uh character, move up because you will be caught upon. See, I have two mics. There's a reason. Okay, so let's just first uh do a quick temperature check. Um raise your hand if in 2025 you think you were pleasantly impressed by uh breakthroughs in AI uh for the agentic economy. Raise your hand. — Wow. Okay. Uh, raise your hand if you were disappointed by progress in AI for agent economy. — Disappointed. — Wow. Concentrated speakers uh all raised their hand. I don't know where you guys were on your B cycle like you know hype cycle. [gasps] I know uh you know uh Shaw story. As of April, AI is supposed to be killing us all by 2029. So, — well, we're uh you know, but we will have a beam chain, right? — We will indeed. So, we'll be fine. — Exactly. That's what I thought. So, um Okay. Well, then how about on the crypto side? Um how many were pleasantly surprised by progress on the crypto web 3 blockchain side for a gentic economy? — Raise your hand. Okay. What about disappointment? — Okay. What about I don't even know what's the point of blockchain in a gentic economy. Raise your hand. — H raise your hand if you're too embarrassed to raise your hand. — Okay, good. [gasps] — Right. Well, um, second, just a temperature check. Uh, elephant in the room. So, who do you think who's Okay. Well, um, who is working on agentic economy, whether you're working on the agent side or you think that you're working at an API company, but you're working for the future of agentic economy. So, raise your hand if you think yourself falls into those categories. How many of you have built an agent? used the an agent in the past 24 hours, in the past hour, in the past five minutes? — How many of you have uh built or used an agent that is not um just a wrapper around an API to a proprietary service? — Okay, good. — Yeah. Okay. Well, so I um temperature checked a little bit about the elephant in the room question. Well actually let me ask those who raise your hand thinking that there's elephant uh okay raise your hand if you think that there are elephant in the room as in clearly big problems uh that no one seems to be talking about or talking about enough or no one seems to be working on them or the things that you come across does not work. So what are like if you see if you think that within agentic economy there's clearly things that you know that is un uh you know we don't have enough mind share and attention devoted to raise your hand. Okay, you have to keep — Yeah, elephants are herbivores. — That's a good starting point. Uh, but sorry, we will have to ask you to raise your hand slightly uh longer because I'm going to call on uh three friends um who raised their hand. Okay, I'll start with Shaw. Oh, okay. Do you need one of my mics? — Oh, it works when you turn it on apparently. Uh, cool. What was the

### Segment 73 (375:00 - 380:00) [6:15:00]

question? — I love how Sha just raised his hand and then figure out the question later. — There's like a big herbivorous mammal in the room that has a trunk. Yes. — So, what's the elephant in the room? — Okay. Yeah. Right. Um I think that a lot of people talk about agent coordination and like, oh, we're going to have all these agents coordinating and I'm like, show me one. Just show me one doing what I wanted to do. Just show me one agent that does what I wanted to do every time. And then let's talk about the marketplace. I think the elephant in the room is that 99% of the people are trying to sell you something that hasn't built yet, which is uh a working agent. Uh, and what we have is marketplaces for working agents and communication platforms and payment platforms and everything except a working agent. So, — who's going to build that? — So, who what's not working today for you? What's your definition of working agent? — I think the problem is that we're trying to make the agents do things that are out of distribution of what they've been trained on. We're trying to make them do things like trade and predict markets and decide on things. And that is all uh stuff that they have been trained on Reddit. So, they have like the Reddit version of how you bake a cake, but they've never actually baked a cake. Um, and I think that we're all basically laring until somebody actually builds a full end toend training pipeline to improve the agents. They don't have the data in their distribution. It's totally a LAR. — All right. That's a really good start. Okay. Well, our friend over here, — you can have mine. — Hi guys, I'm Tim. All right. Um, the other elephant is that inference is really expensive. Even the tiny LLM inference calls that we're trying to somehow sell with X42 services, you know, every time a decision is made, — it's more than those pennies stack up really fast and a lot of people there are people in the room right now who are working on awesome DPIN for us to like bring down the cost of intelligence. But uh yeah, there's a lot of subsidization happening right now. What's your inference bill like per whatever unit of time? — I am an A16Zbacked startup and I got a lot of free inference from all the big cloud providers. So, thank you to all of them. Um, and that's how I'm surviving. — Well, uh what do you think is the right uh like what's the multiple of uh you know what you think is acceptable right now? Um, when I really when I think about just well it needs to be a fraction of X42, right? Every decision needs to be essentially 10% or less of what an X42 transaction cost would be. — That would be an amazing I can make a decision and then take a fee for it. But also now the service itself, the service might be more expensive than that. That's fine. Charge 10 cents for doing a forecast. That's fine. But the actual should I fulfill or not? Even that binary decision, if I have an LLM wrapper on it, that needs to be really tiny. — Great. Well, um, any other elephants? Okay. Well, right. Right there. Next door. — Politics. So, — the duck vision. Yes, the duck vision and what we're working on. We need to build with the right technology and we're doing that. But let's think about permission discovery for example. But the reality is that there are market forces that are that could directly bring to having like a centralized GPT store that will intermediate all the agents that we will use. And now like regulation is not uh is not cool, right? We to damage US companies. And the risk is that while here we build the right stuff since things will happen very fast, the wrong decision or the missing decision of uh like the global governance in general would just bring humanity in practice in the wrong direction. give me uh an example of something that you have experienced or uh you know where politics become like the biggest blocker to getting uh it done for in this context — I'm a strong supporter of uh like uh pushing players to have interoperability they run a marketplace I think that we should disjoint the quality of the front end with the data set so I think that we should like force who owns huge data sets to make them public otherwise they will become the one place to go — and um so well I can make examples on specific regulation on I don't want to mention company I'm referring to I basically want to avoid the agent app store and I think that we are doing the right stuff in this room but frankly I'm not sure if this is enough that's I think my elephant in the room — that's great well the new game rule is that whoever has the mic uh needs to run no no like you need to run it to the next person that gets called on this. Got to do some exercise. You've been sitting here all day. I've watching. I've been watching. — Okay. Well, uh, Andrew Miller, uh, gets, uh, like there. Okay. Well, sorry. Uh, the mic went to, uh

### Segment 74 (380:00 - 385:00) [6:20:00]

Mike. — Hey. Uh, I love this question. So, uh, Devin, um, yeah, probably two elephants in the room. One, uh, settlement's complicated, right? So as we move to this agent economy, how can agents take best really have a best practice of settlement right hitting the RPC settlement mpool block? — I'm going to ask you a half troll question settlement because you know that public blockchains today do not meet legal definition for settlement FYI. — Okay, fair point. Uh but landing in a block, right, in consensus. So that's one because it's a complex dark forest, right? So how can agents just take advantage of best practice for that — being in a block and then secondly um I think another elephant is how can we make agents fun right celebrate arts culture we have an artist here seven here who's uh has an incredible uh art stack running in a te generating information diet pictograms that's fun that should be celebrated as well — wow yeah let's keep the clapping going okay well choose our favorite Mike. — All right, Andrew. — All right. Uh, reputation systems never work. They really favor incumbents a lot as well. So, I think we're kind of expecting reputation systems to bail us out. And I think that's an elephant in the room. Also, open source agents would be great, but I think all of the good agents are not going to be open source. And so, we have to get trustlessness somehow. Uh, anyway, — well, how do you like I think you haven't mentioned the elephant though. Okay. Open source model. People don't want to open source. So what — models or the wrapper codes? Yeah. I mean I think te sandboxes is kind of the way but — you're biased. — Yeah. — Yep. Uh why? — Oh, because I I think you can have open- source agents go into the sandbox and evaluate the closed source agents for safety. — Oo. Okay. Now we have an interesting elephant. Okay. Well, uh, pick the next. — Yeah, my elephant is about bias in the models. You know, you have these closed source models or open weight models, but we don't have a foundational models that are open that we know in which data they've been trained on. So all these models are biased and we are delegating more and more of our day-to-day lives onto AI and Asians and we don't know what these models were trained on and we are delegating these decisions to things that we don't know how they're going to work if they represent our best interest or not. — Yeah. Uh, speaking of AI bias, uh, have you guys seen the, uh, the latest stuff about how Grock seems to think that Elon has like a really great body [snorts] — and — well, we we should — you need to connect this to your answer response to that question. I mean, we, you know, we have to know how to trust the AIS, especially if we're going to trust them for uh, you know, things like determining the future of our culture, which like lots of people seem to be like jumping headong into. So, what's the, you know, — Yeah, exactly. I only we only have like five labs doing all the foundational the big foundational models and we depend on them completely on how they train the models and what data they decide that goes in and then the finetuning as well. They do a lots of of the post training. Lots of biases there in there as well. — Yeah. Well, uh this we're gonna take one uh more. So, pick someone that Okay. — Thank you. — Proximity. I'm telling you, latency is all you need. — Yeah. I think um agent should operate on a trust minimized. So basically whatever your code is running inside the TE or not the smart contract should be the gate uh to actually stop those agents. So we can actually give them access to do certain actions. I talk about DeFi agent for example they can move your funds but they can never withdraw it. They can never use certain action uh actions. So basically trust minimize means they can do certain actions but we should always enforce inside the smart contracts on chain. So I'm think if this is something that we need to focus more on trust minimize on instead of trust list that would be great. So would love to know what you guys think. — I'd like to hear Eric Cifphant. He's been working on this stuff a lot. So — yes, Eric uh please grab the mic. — Um I think that I think the elephant in the room right now — uh closer to the mic please. I think if the elephant in the room right now is that if we don't end up with an alternative to these verticalized giant agent conversational LLM interfaces

### Segment 75 (385:00 - 390:00) [6:25:00]

we're going to see the deterioration of the open web as it exists. I think that if we don't have open source and open standards that can compete with the full stack nature of the big AI labs and the big interfaces that all of us use every day, the open web is just going to disappear. — That's the perfect segue to uh the start of uh the town hall panel. Um here's our um uh friendly co-hosts. So let me start with a question that may uh be kind of like uh you know maybe someone I'm the dumbest person always in the room to ask this question. So what's the role of blockchain in agentic economy? — Let's just start there or Ethereum. Use whichever you prefer. No, — you're the guy in the agency. Yeah. Uh I mean uh for sure payments and like microp payments, the flexibility of payments, how cheap they are and now that we have scalability on blockchains, we can actually do these nano payments. So that's one where you set value like transfer like because agentic economy is about like uh value transfer between like different entities. Uh so that's one and then the other one is uh everything around trust like yeah I agree that like uh it is not clear like uh what role reputation will have and like I disagree with Andrew that like reputation system never work like we actually use them like in web two like they are widespread and for specific like vertical use cases we use them like every day I think uh they'll take some form here I'm not sure what ERCA 2004 is not like a reputation uh system. It's just like standardizing like some data structure that then people can use to like build like uh reputation in like different uh applications like different verticals. Uh yeah and the other one is like um kind of uh hard trust based on cryptography as people mentioned — and so using the chain for like uh as a verifier basically — using the chain as a verifier. Cool. Okay. What about you? — I mean the other big one is of course that like onchain applications, you know, what I would what I call onchain games are like exactly the things that often it makes sense for agents to participate in. — Why onchain games? — Um well I'm using game here and you know the expansive sense in which you know people use the word game theory right. So uh a market is a game right and uh the thing I always say is that actually we've had AI blockchain intersection use cases since 2015 which is uh um automated uh bots trading on onchain dexes and uh to me that was just the earliest member in a category that will grow larger and uh I expect to see uh you know agents participating in prediction markets. I expect to see agents uh participating in governance of DAOs. I uh expect to see yeah them participating in all kinds of these structures and in some cases even ideas that were possible in theory but impractical because of human factors until now will now become practical. — I'm going to uh dive a little bit into your answers. M — so you seem to imply that um onchain games onchain markets on onchain uh like uh essentially uh interactions — what are the unique value proposition — that blockchains provides that are more suitable for agents than the human factor that you mentioned and glossed over. Yeah, I mean the crypto environment I think is uh one that is I mean actually very na very naturally suited for a more kind of anonymous way of uh deal doing things right where you know you don't need to know anything about uh someone's identity as a precondition of uh being able to trust them and that's something where agents can uh very naturally slot themselves in Um also in anything computerized will inevitably go up to very high levels of technical efficiency to the point where the bottleneck becomes how many decisions can a human make. And so if you have something like an agent that is um you know uh fine-tuned off of you and like contains some like reasonably good encoding of

### Segment 76 (390:00 - 395:00) [6:30:00]

your goals and preferences then it will be able to make a much higher volume of uh decisions on your behalf and uh I mean the reason why like blockchains and not like running these things on top of uh centralized systems is basically that there's a large intersection between applications that are like that and uh applications where there's a need for a very diverse group of people to have a high degree of trust in them. — Do you have anything to add on this elaboration? Yeah, maybe like I want to bring to the conversation like u there is actually like people uh using a F4 to like run some experiments of like new type of games uh with agent like Shaw and Marco today presented like this game where like it's effectively like um a simulation of like prediction market but with a clock that goes much faster. So like the market gets resolved in the game. Um yeah, maybe like one question is like um could this like could the chain also uh serve as like kind of like a data generator like somehow of like uh — well a chain is nothing without its users, right? So you know the users are the data generator. — What if the users are like some agent personas themselves? — Then they can be data generators, right? I mean, when's the last time you looked at AI generated data? I mean, for me, it was definitely today. — Yep. Same. — I think this brings to like um one of the uh questions I've always had is it seems that uh well, Ethereum blockchain have a lot of like uh essentially post uh post settlement like everything is public — but of course you can also post proofs. what kind of uh but then there's different kind of camps where if there's more data uh like that gets posted on chain perhaps it may be easier to democratize the ability um on uh the training inference side — but then on the other hand potentially you just end up um having uh essentially um privacy concerns amongst many others. M — so again I'm still want to I want to go back to what is the role of the chain and how do we reason about it in this particular context. — I mean the chain is uh like it's basically the place where mutable state uh or shared mutable state uh lives right it's the place where you know objects uh live that get modified and where multiple people could be modifying in them and uh like and care about what their value is, right? And so that includes a lot of financial assets. It also includes lots of other things too. It's not the whole but a part of almost any application. So what uh again like what can blockchain um uniquely enable in the agentic economy aside from like I think that DA you mentioned a lot about like you know uh the practical use cases like payment and also I think uh you two both uh mentioned kind of like as a verifier. Can we think of a couple more that are uniquely enablement uh enabling? — Maybe one area that is interesting for me is like constraint delegation. M — so like uh as automation increases essentially what we are doing with humans is like okay we are delegating part of our agency to like uh this uh like some system that may be autonomous and it's like uh smart contracts can provide the hard rules that like uh the system should abide to and we actually had a couple of projects today like presenting like some products they're building in this space like um yeah — yeah um fits right into you know account abstraction and smart wallets right you know you can have accounts where you have arbitrary sub permissions and uh you know you can hand any one of those keys to a bot or not do that — anything else — or raise your hand if you think you know you have an idea of what blockchains uniquely enable or Ethereum Um, whoever has the mic last or someone else magically produced that mic else I'll have to give mine. uh it's not entirely clear uh what these

### Segment 77 (395:00 - 400:00) [6:35:00]

new standards are providing that are specific to agents rather than to humans. Uh and — I think that so there was an experiment a couple of years ago a few years ago where Google had two AI chatbots talking to each other and eventually they started making their own language and the Google researchers started freaking out and shut it down immediately. Um, I think that the one of the main applications that we're going to see of blockchains that are specific to agents and not humans is probably the creation of synthetic assets that don't make sense to human beings because they're just complex combination, you know, baskets of goods of like all these different things we don't really quite understand. Uh, and uh, that like humans would generally not come up with, but that but that agents would. Um and then one other point I just wanted to make real quick about your earlier question which is about this is called trustless agency. Uh and the issue is that I see a lot of uh assumptions of trust uh this like reputation assumes trust like you're trusting that reputation score generally uh and without you can't have uh real trustlessness without some level of escrow and arbitration. uh you can't act you can't like you can't just immediately pay somebody uh for a service uh because that means you're well you can but then you're trusting them right that's not trustless so if you want trustlessness you need to actually like have something in escrow make sure the service is actually provided and then be able to arbitrate over it uh however you want — yeah I think it's a good point and actually there is a few projects that are building this like credit uh protocols um rating agencies is so yeah — yeah I mean generally I think for like things at those small scales like trying to make it trustless within the scope of one interaction generally doesn't make sense right because like there have been like a decade of people making like fancy protocols where you try to like trade you know coins for data and then there's like even I think theorems that that's like not really possible unless you accept some pretty weird limitations and you know the actual reality right is that Like if you uh like if it costs $1 to sort of set up an identity, then like you will be able to just uh trust them in like one cent at a time, right? And you just like do interactions back and forth and they can only cheat you to the extent of one interaction and like that's fine. The one case in which that's not true of course is privacy, right? Because if uh if an interaction involves trusting uh something with your private data, then like in even in one round it can totally reveal everything. And like the other problem is like you can't cryptographically prove that you did not leak data, right? So that's you know exactly what going beyond cryptoeconomics and know into cryptography is for — great point. We're get we're going to get to that um here. Yeah. — Economics. You know, I hope I can phrase this correctly, but I just want to follow up on what Eric already talked about that we have this unique moment in time where the incumbents with all of their billions and near trillion now are building these state-of-the-art models that we all rely on that are biased in certain ways — and that we can't how will we really enable like a public gift to humanity if it's only under like one, two or three people's control. — And it feels like we have done this incredible job over the last like almost 20 years but 15 years of decentralizing everything we can and coordinating hundreds of billions of dollars at a time. Can we enable the agents to help us train a great work a actual public model that is always state-of-the-art, always building data pipelines, always doing these things because the agents are self-incentivized to do so using hopefully what we're building here. — Yeah. — Uh I'm going to um call upon our friend from Google. Um so like one of the labs. Uh so what's between uh everyone us here and um the state-of-the-art models capital compute data and anything else and uh what can blockchain um and what you know we do here matter. — Yeah. Wow. Big question. Yeah. I would say um so when it comes to these blockchains right you have public providence right public permissionless public records right that can be inspected verified uh checked by anyone anywhere right and I think that's the critical mechanism where as long as these layer ones and al like can keep that principle that matters right as a foundation to build upon I would also argue I mean all the

### Segment 78 (400:00 - 405:00) [6:40:00]

infrastructure bits you mentioned uh and similar to previous panels here uh it goes down back to infrastructure right you know how do you trust that VM right that's running that container how do you trust the source of that VM the supply chain that built the VM the silicon the RAM I mean how far do you want to go uh down that path and there's a lot of work being done you know with um teams and trustless TEES and uh you know people taking a uh a hardware perspective right from NFAB up to build trust at the lowest of levels so I think that's where it to start and ideally that's where also I think the industry should be investing right in terms of the academic research and uh the the designs because it it really has to I think start from that primitive that physical primitive — so we need to first build chips before we solve the model problem — sorry I'm trolling you a little bit um but I do think that yeah what uh like B what do you think like what's between us and the trillions billions. — Yeah. I mean, I think uh my kind of instill low confidence but current view on that situation is uh that uh the less realistic thing is to try to beat the big centralized guys on their own turf and make a model that's sort of better at all of the metrics in an absolute sense. I think uh the big advantage that uh open models have by virtue of being open is that you can fine-tune them, right? And uh that you know they're not going to win on general purpose metrics, but you can always make a version that's uh specialized for any one application or set of use cases that you care about. And I feel like the ecosystem is like not like it still could do more to really uh lean hard um into that and uh like take advantage of uh of that as a primary feature. um uh to the because like an agent uh like it should an a like an individual agent should not just be a prompt string right like you can even with you know Laura like you can make it a fine tune and that's still like what some number of megabytes right you can uh go much further and customize and uh like I think that's uh the way in which you can make things that are open and that actually you know serve individual and local use case as well, right? Because like the native advantage of decentralization always is that it's better at serving like individual local needs that are inherently not scalable, right? And so we should try to do that — and diversity and that's um — yeah that's why we're working on these standards to like uh we hope like by agentic economy we hope to like leverage the diversity of like uh the ecosystem. — Okay. Well, um this is a good segue into uh Silvallic, you wrote a um blog post at some point last year called the glue and co-processor — architecture — and uh it was inspired by one of the conversations with Andrew I think um and many people here actually review that post um as well. What's interesting about it also like you know Mike's um sorry uh Devon's uh like uh um uh answer just now around the silicons and also the foundations of trust um kind of lead into like basically what's the like five year five to 10 years down the road. — Yeah. — Um what would be the future architecture of compute look like — for a gentic economy? what should happen where — like how do we reason about local first? like you know essentially um yeah the substrate? — Yeah. I mean the thing that like frustrates me that I think is like an actually hard problem right is that I think like local first is the yeah currently the only thing that I would truly trust right um and uh but then the problem is that like the fundamental economics of it are that uh each individual's usage is inherently spiky and uh if you have hardware that's enough to meet the spiky part of your needs then it's going to effectively be idle during the non-spiky part of your needs. And so it becomes very uh not cost effective because you're like you're overpaying quite a bit on uh on compute. And then in some sense you're overpaying for electricity both because like raw because of raw costs and because like you're if you have like a laptop or a phone or whatever, you're not just paying for the electricity, you're also paying for the annoyance of like charging the battery, right? and uh and so you're sort of at a big disadvantage compared with the remote stuff. And uh it would be nice to try to

### Segment 79 (405:00 - 410:00) [6:45:00]

find some kind of path around that is not trust dependent or that is as uh minimally trust dependent as possible. And uh I mean so far I think this is still like this is still an unsolved problem, right? I mean, my kind of medium-term view of compute is like my uh like my controversial kind of the thing I believe that few other people believe is that I think there's a significant chance that laptops will be dead in 5 years. And uh the reason why is basically that uh it's just obvious that the correct thing to do is to kind of disagregate compute from UI and uh like when you have one compute then you have all kinds of UI form factors that are just multiplying right and like we have originally you had phone and uh and laptop and originally had laptop then you had phone and then now you have glasses you also have like watch and then eventually you're going to have BCI and uh like it just makes natural sense to kind of break that apart and then break out the uh compute piece and uh ideally like at least figure out like the low-level stack and the operating system so that uh like who people who wants to do that stuff locally or even in any format that they personally trust are able to do that right so like your decision of who to trust is like fully yours on the basis of who to trust and it's not like bundled with you know, choice of which applications you use. Um, but like this is I mean like I'm not giving an answer. I feel like I'm giving you like a bunch of unsolved like really uh challenging and uh interesting design problems here. — How would you even like you said is like a big unsolved problem. — How would you even you know start going about solving something like this? Yeah. How do we think about it? Um I think it's like one of those like I would sort of start use case by use case right like bas because like unless you build something that serves even a small amount of concrete needs fairly early on like you don't have that feedback where you know whether you're actually solving like valuable problems right and uh you know like basically start with like some subset of uh applications that actually can be done in a more trust minimized way and then you have like you know the spinning wheel like you have an ecosystem and something that can then like actually expand from there and it has the incentive to keep figuring those next steps out — personally which app would you if you have unlimited time just like to do uh some fun things or solve hard problems which application would you start with solving this problem — I don't believe the uh the answer is application level I think the answer there's like a lower level and of course you know there is the kind of galaxy brain take that everything is an application which uh you know can be debated but I don't want — I believe so but — but I don't want to extend this panel to three hours um but I mean I think at the operating system it's uh like I've talked about how you know like you want to like with like in crypto you know I believe in the sort of thinner DAP thicker wallet vision because I believe that there should be more sophisticated things that are incent explicitly incentive aligned with the user and that are able to protect the user from things that are happening at the other ends. And so I equivalently like I would love to see more work being done on operating systems and on figuring out kind of the pipe the piping to uh make uh that stuff more open-source first, more local first and uh like where everything is at the very least like very choose your own hardware friendly. Um and uh um so and something that can like replicate the functionality of like basically you know the current entire uh stack of uh centralized trusted things that people depend on. Um I mean if you want like examples like live translation right like uh that is something that uh a lot of people depend on but and uh you know you depend on more the further you are from you know the core of you know the sort of English speakaking you know great empires and uh the and so it's a big need but at the same time it's like if you look at encrypted messaging right the encrypted messaging is uh some like it provides privacy properties that people really value and so you think that when you're using signal the message is just between you and the recipient but then you don't know if the recipient is uh using you know like some

### Segment 80 (410:00 - 415:00) [6:50:00]

centralized cloudy AI thing to translate the messages and potentially like you need some way to translate the messages as well right and like by default that's a big trust like that's like that violates the stated guarantee right and so you won't like that's like If there's like one thing that uh I would um like if there's one thing that's like low difficulty to make trustless that I would make trustless within you know the AI world like translation just makes obvious sense. Um and then from there you just kind of keep on you know increasing the uh set of capabilities. — That actually uh brought us to a really like another interesting segue to the what's the right privacy model? But — in your minds the privacy model for agentic economy? Um, one of this question is a shout out to uh my teammate Tessa who uh listened to the next door commerce uh uh like panel uh where you had an interesting discussion with the tour guy um not a tour guide but tour guide um on onication of um essentially um — mean Ethereum and everything. — Yes, exactly. So yeah and so like how what would yeah a privacy model be like in agentic context? — I mean I think uh you know the first thing about privacy is that privacy is not a feature. Privacy is a hygiene. Um and so the uh you know the correct metaphor to think about is not you know these are new capabilities that you're adding but rather these are data leaks that you are no longer leaking. Um and uh like I think you wants to start with you know like that kind of principle and uh apply it to the NG& techch world right so what are like things that you do not want to be revealing what is information about you that and your activity that if known could lead to you being exploited and uh like how do we find ways to minimize that right and I think there's like a couple of different ways to approach the problem. There is the level of uh approaching the uh problem like there's a sort of monolithic perspective where like you assume a user is aided by a pile of stuff and that pile of stuff is uh trusted but then it interacts with external stuff that's untrusted and you want the pile of stuff to access your data so that it can help you know be a functional cyborg. uh but uh you do not want but then you want to like give instructions to your pile of stuff to treat your data with care in a way that it does not expose it to other people's pile of stuff in a compromising way. Um and that which could mean you know like exp like exposing computations on it in an uncompromising way. That's kind of the whole point of know cryptography right. Um but the other approach that you can take is like realizing that even your local pile of stuff is sort of not fully trusted and like you need like multiple layers and uh like you have a thing that is then giving permissions to another thing and that thing is giving permissions to a third thing like effectively how you know you have like you have a browser and then that browser is giving permissions to websites like it's the same kind of hierarchy. Um, who here works on privacy? Raise your hand. — Um, I can't uh see your face, but please uh you know hand whoever raise their hands a mic as in like — if you can stay succinctly what do you think is the right — like how do you think about the privacy model like uh for agentic economy Does someone have a mic? I can't really see but Oh, yes, please. Which one? — Can I go? — Oh, yes. — Yes. So, I'm working on AI agents in decentralized cloud computing. We have a network where we tap into it to create the inference. Um and what we see is that we use CES for the inference because uh the cloud computing it's hard to scale if we use CK proofs for example it's too slow for the inference and so we use the inference we do the inference on of chain with proofs on chain of the integrity of the models. Uh so how we can scale privacy when CK

### Segment 81 (415:00 - 420:00) [6:55:00]

proofs are not there yet in terms of uh capabilities and and speed. — that's a question I think. Um so I'm going to call on uh Andrew Miller who was playing it with his phone so he probably didn't listen to the question. So, how do we uh scale privacy uh like um uh for agent context when ZK seems to not scale? And I think this is I love a banger answer from uh I request a banger worthy one from Andrew Miller — is the answer TEES. I was listening to you know the explanation you're already using TEEES. There's lots of them. All of the cloud companies have them. while the Nvidia is already using it. Is there um more to the scale question? — Okay. Well, perhaps uh like can you give a quick breakdown of the trade-off space of what options what toolkits uh you know we have access to today and what you know how to reason about it today and in the future. — I mean really it's just um I mean the trend of the it's hard because there's so little margin, right? Everyone wants the fastest hardware and so there's not really a lot of room to add overhead. No one wants to go slower and then try to compete by having more security as um you know cutting against that budget. And so something like maybe regulations or just communitywide insistence that this really matters. Um I really love the stage one stage two thing. Um, I think that's like a really cool innovation in just pushing public discourse forward because no one wants to make their roll-ups go fast, you know, slower unless they have to, right? So like insisting that you have these extra features in hardening comes from a different place than what you normally have as just productwise competition. So I think we need more things like that to just insist on them. All of the security and engineer people at the hardware companies um really want to do this but then their product teams you know say well we need we can only do what our customers demand for us to do. So the you have to find where's the leading edge to push that — like uh doesn't this uh kind of well it seems to be motivating uh a stage one stage two not only for decentralization but also for privacy and security um — both for the privacy and security to be there I think yeah and Vatell was going — right I'm so tempted to uh ask you both like what's your stage one and what's your stage two but we'll get there but okay respond Yeah, I mean I think uh in addition to TE is the other strategy that sometimes gets underrated is anonymization, right? Like if you can't hide the contents of the request, then the second best thing is to hide the uh source of the request and sort of hide among the sea of uh all of the other people asking things around the same time. And uh this is something that is uh I think very doable like you could basically have you know the uh equivalent of a mixn net but like basically specialized for like routing requests to uh AI models and like that doesn't work for every use case and like you'll probably have to give up on your custom prompts if you want an anonymity set but uh it's still a thing that I think is worth building. Um yeah um I guess uh David one question for you as I think you know you're in the Ethereum Foundation's new AI team and I think you encounter a lot of people who think about AI and agents from uh totally different angles and people who also think about things like privacy and even blockchains from different angles. What are some big like divergences that you see in uh how different parts of uh the space think about the problem and like who is not talking to each other, who should be talking to each other like you know. — Yeah. Uh this is a great question and it's part of what we've tried to do with this event as well like um since uh we started the team like we wanted to like focus on like two things. So we call the team DAI all lower case D and where D like stands for like um decentralized uh coordination of AI which is what we're trying to do with this protocol and standards we discussed today. The other one is like more defensive AI like uh we are collaborating with the protocol security team at the AF and we are also like — doing some initial exploration on like okay how can like uh we help like uh AI security with like some of uh the tools we develop in blockchain. Um yeah, in terms of divergence like I feel that uh our focus since starting the team has actually like been mostly on this decentralized coordination simply

### Segment 82 (420:00 - 425:00) [7:00:00]

because like unexpectedly after we launched uh ERCA 2004 like there was kind of like huge excitement in the community and like we had like hundreds of people kind of like joining like most many of the people here it's like community that got formed in the past like two three months. Uh so as you see like there there's not too many people working on uh privacy in uh the room. I think that's kind of like something we've tried to do a little bit with this event like uh we uh of course like Andrew uh has been collaborating with us. There were like two three other projects that are working on like uh many like uh TE applied to like uh agents and AI. So effectively I think those communities I feel we are trying to already like uh catalyze convergence into those but uh we could do more uh and um I feel that actually like this area of like uh privacy in these protocols right not only 804 but also x42 may be like the focus of the next wave like to me kind of season one was like this past three months in which like this community got together and uh we kind of lifting these protocols off the ground and then season two is like kind of okay now we have these standards uh what can be built on top of them to make them like more usable for developers more extensible and then like more private like um in the case of X42 like one uh big point is also like the centralization because there is centralization vectors kind of similar to what we've seen in block building uh as payments will start flowing and there is a role of the facilitator which is kind of similar to uh the block builder. Um so yeah um this is going to be the focus. Yeah. — So I want to like uh I love to ask a road map question. Mhm. — So, um like why did you start with um AO4 um instead of like you know of all of the problems you could solve in like uh agentic economy like what's uh why start with here and what is coming next — and I still want to ask about the stage one stage two of privacy uh straw man uh you know that's another road mapap question Yeah. Uh we started uh thinking about this like um in the early summer and at that uh time like uh the X42 protocol was uh already like going live like not many people knew it. We were talking about it in some events we organized at the Ethereum house. Um uh but uh like essentially like I felt that uh the microp payment or nanop payment side of like the agentic econ essentially okay what is the agentic economy is like um uh creating like the rails for like uh different entities and even different services to like interact with each other and uh pay each other hopefully in a trustless way. So like payments was kind of already solved in my mind like I felt that okay it's just the v 0ero of x42 but this protocol like if it evolves in the right ways can solve uh the payment side of things and then like um the missing parts where like uh okay if now you have like this universe of services right like the cool thing about the agentic economy is that you can have like high frequency uh economy similar to like what you have in high frequency trading like because if have like automated agents there is going to be a ton of interactions. So then microp payments is solved then like you need to discover all these microservices and uh somehow you need to trust them. So I felt like uh with Marco which is my co-author original like coming up with this idea uh we felt that this is kind of the missing piece and uh it's been pretty exciting because like now that like both uh X42 and 804 are starting to be used. People realize that there is actually great complimentarity between these two standards. But I think it's just the beginning like we need to do more build more on top of these standards and probably other things. — So what's to come on your road map? — Yeah. So we are going on mainet with ERCA 2004 in a couple of weeks. So that's the next milestone. Uh yeah and then after this um uh we want to like uh continue working with the community on two areas. One is we already have projects building public goods uh for these protocols like um alt layer um built 804 scan. Um some other

### Segment 83 (425:00 - 430:00) [7:05:00]

people are thinking about like uh um like SDKs to like uh make uh developers life easier to interact with these protocols. So on one end we need to like grow the uh infrastructure uh on top of these uh standards and then on the other end I think we need to continue like experimenting like to me the way I see this evolving is uh almost like it's a new ecosystem the mass of things we experiment with on it needs to grow until the point that there is a critical mass and like someone can build like a multi- aent system that really leverages like spe specialization and can give an output that is kind of like very good and better than what can be done outside of this. So I feel the next few steps will be around like um uh curating these ecosystem so that we can produce those killer apps kind of similar to what happened in DeFi in the past. What's your wish list of other ERC's and other uh standards um or uh protocols that you see that maybe not yet on your main road map but you would like to see happening? — So one that we're already working with um uh 804 already has like a validation registry which is the third part of 804 that we don't talk about uh that needs to be more fleshed out. Uh it's the part that um we started but we are collaborating with Andrew with Justin from Spar Sparity. Um so that part uh is definitely like something that we want to standardize but besides that like not everything needs to be a standard right uh some things will be kind of like offchain uh systems. Another big one which is adjacent to 804 is like um we need to decentralize uh X42 uh discovery right now is like um basically is still the initial PC of X42 and it used to be very centralized. There was just like one party like the Coinbase facilitator now that we already have like a dozen facilitators but like the there is some protocol design there to do and so it's quite interesting. uh we don't have solutions but it's an interesting problem to tackle. — Cool. What about you V? — What other ERC is — um — or EIP or anything that in terms of community um bottom up um standards or shared best practices. Um yeah. — H — or maybe I have a different question. Can I ask? — Sure. — Yeah. So like um you have experience also like working with communities of course like in the past of Ethereum there were multiple times where like um people get really excited about like some new development or new direction like what uh tips do you have like for us and for this community like uh — communities that are excited in a direction as opposed to communities that are excited directionlessly which exist too. Um yeah, it's I mean it's yeah it's a good like very yeah abstract question. I mean both of them like I'm trying to think more um yeah what's uh I don't know my mind is blank. Yeah, — I think the Oh, — it seems fair time, but — yeah, it looks like uh we're getting kicked off the stage officially. Um so like if you are interested in accelerating the community um uh directionless community looking for directions or direction uh you know uh you think you have a direction but you want to accelerate it. Um come hang out at the pirate ship. find me after of DMU the coordinate. That's how we accelerate. — Cool. And I want to thank everyone uh for today. It was a great event and uh thanks Vidalik and Tina for joining us. Thanks. — Yeah, thank you everyone for uh on demand participation and uh let's chat more um tonight if you're around. Um I think these are really important questions. — Heat.

### Segment 84 (430:00 - 430:00) [7:10:00]

Heat.