Agent-first workflows from prompt to production

— All right, it is great to see you all. I think most of you are here for the air conditioning, but nonetheless — this will be good for you regardless. So, welcome. That was a great dev keynote earlier if you all got to watch being able to see how much all of us could create software from prompts, from our voice, from context. How amazing is that with anti-gravity 2. 0, with AI studio? This is amazing time to build. So, we're going to have some fun today doing a lot with this. Now, first we just built this game in uh AI studio. If you saw the dev keynote earlier, you saw that they were building games, doing all kinds of stuff. Let's have some fun. So, when you build a game like this, you might click publish. You might deploy it to something like Cloud Run in Firebase. And hey, no credit card required as you start shipping your first two applications, which is amazing. We all love day one. Day one is amazing. You ship it, you feel good. But then there's day two. There's real traffic. There's things that go wrong. There's edge cases you haven't considered. So, what I want to do though is, you know, I want to introduce first my friends Christina and then uh Denise are here to help me out. They've built some amazing things. Yes. [cheering] They're both amazing. What I want you to do is take out your phone, scan this QR code, and let's play this game. Seriously. I want to see if it can survive IO traffic. I have my doubts. Uh Christina's job depends on this. There's no stakes here. It'll be fine. So, what I want to see, let's try to play the game for a minute and see if you all can break it with load or it'll be completely fine and your job will be safe. Either way, it's not a big deal. So, we're going to go here. We'll look at the leaderboard, I think. We'll see Oh, yeah. Let's I'm going to go. You're unbelievable at the game, Denise. So, let's see. Let's see if I could beat this. It's looking good so far. — You only need to survive 10 seconds. — How's the leaderboard doing? Okay. — Oh. Okay. All right, let's see how the leaderboard's is — check out the leaderboard, see if we're holding on. Huh. Uh-oh. Well, I'd like to thank Christina for being here at Google. Uh we'll — give you your paycheck at the back. Uh-oh. Oh. We'll pay out your vacation. We're very humane with that. No, it's great so far, but this stuff happens, right? Things can break, day two happens, load happens, things like that. It's okay. So, how do we have to think about a little more than that? Cuz if you ask most developers, where are you spending your time with AI? Guess what you're not doing. Day two stuff. Deployment, monitoring, fixing. Most devs are like, "Hey, I'm using it to build document code. " That's awesome. But not enough of us yet are making this part of our day two journey. So, how do we think about that a little bit more? How do we actually take advantage of what AI can do for us on day two? So, over these next 30-ish minutes or so, we're going to show three agent-first workflows. These are going to take these prototypes and turn them into something more reliable, more production grade. We're going to use antigravity, Google Cloud, a number of components. Let's do something to make any of these fun vibe-coded apps something more legit. So, we want to start with agent-first debugging. I mean, some of you might like debugging. It's fine. You're probably fun at parties because you talk about this a lot. But Denise, can you set the stage for us here? What's it look like to do real agent-first debugging? Yeah, so I mean, when something crashes, most of the time you're going to have to go through all of these logs, and who really wants to do that? Cuz I know I don't. You know, you got to go look at your configurations, see if you made a mistake. You go look at the performance metrics to see if there's any issues with latency and everything, but nobody wants to do that. You know, so what if we took our agent, our agentic developer tool, in our case we're going to use antigravity, and what if we took that and we let it do the job for us? Right? Why do all of that manual task ourselves? So, it crashed, and we want to know what happened. So, let I am going to use, because I'm kind of lazy, I don't want to actually type it. I'm going to use the voice input and ask, "Find out what's wrong with Dino Quest. " So, going to kick it off, and notice how I do not actually tell it, "Where is my logs? Where is my Dino Quest service? " I was able to do it all with just telling it what to do. And the reason why I was able to do that is because we have the Google-managed MCP servers. To be able to safely connect our agentic tool to MCPs and our Cloud Run and our services. And so, by doing that, and as you can see, there's lots and lots of them available, we don't have to juggle with API keys. go into the console and go click around and do all of that stuff. All of our data and our services from Google Cloud is available for the agent. So, let's go back to Any Gravity and see Look at it right there. So, as I was talking, all of that happened, right? Like, would you normally be able to do that if you were manually debugging? Nope. So, it told me that this is actually caused by a 503 service

unavailable. I think all of you putting it there put a lot of the issues onto a load. So, going on, um we know what the issue is now. We want to actually fix the problem. So, I'm going to ask Any Gravity again. Okay. "Can you look at the code for the Dino Quest game and fix what caused the issue? " So, here, send it through, and it is going to go and work and do all of that. But, as it's going, Richard, why don't you tell us about the developer knowledge MCP? Because I think that's a great MCP for people to use. Yeah, I mean, you mentioned Look, there's over 50 of those managed MCP servers you can use today from everything from BigQuery to Cloud Run. One of them, which everybody should use, is called the developer knowledge API or developer knowledge MCP. What this is an MCP server that fronts over a dozen different Google Doc bases, from AI to Cloud to Firebase to all these different components. So, you can just say, "Hey, I need something. " And it goes and searches any of our docs and pulls back the latest version into your coding tool. So, all of a sudden, you're not guessing, "Am I using the wrong version, the right version? " You're getting a 8 to 12-hour at the latest snapshot of the documentation into your agentic tooling. So, doesn't matter if you're using Cloud Code, you're using Code X, you're using Anti-gravity, pull the latest of Google documentation into your tool and make it easy to use. Yeah, I mean, that's very helpful because one of the problems is always using outdated date like outdated information. And then your agent's not going to actually do what you want it to do. — All right, so let's go on and looks like Anti-gravity has finished and it said that the issue was actually in our leaderboard. And it went ahead and made the code changes for us over here in the main. py file. I didn't have to actually do anything myself. I just let it go. And so, typically what we would want to do is go and commit it to GitHub and we're going to redeploy and Anti-gravity can do this for me. But because we have a little thing a few things up our sleeve later on, I'm not going to actually deploy it now so that we can show you something better later. I like it. — Christina. Sounds good. I know here in Google we're all about transparency, right? And things like this happen, we wanted to make it out for people, right? So, for those of you who kind of suffer through the game, we want to give you some treats back in the game. And as you know, um what we have right here is Data Agent Kids. Data Agent Kids gives me some kind of visibility into my data source. The reason why I talk about data source is because I live coded this game 3:00 a. m. in the morning. Well done. You're a trooper. Right. This is where my dark circle came from. It was it got very fuzzy at the end. I kind of forgot where I store everything. And this is where this came very helpful because I was able to kind of figure out where things are without me needing me to actually go in and figure out things out. So, what I'm going to do here, I have the data agent kit installed in my antigravity right here. And um, let's go ahead and then start to head over to our antigravity. Right. So, what I'm going to do is I'm going to ask our antigravity with data agent kit to look through all my data sources. And as a treat for you, as you can see all the document databases that I have cuz I've been live coding with AI Studio and AI Studio directly uses most of the time it goes to use Firestore as my database and it's a document database. And to update it is not as simple as, "Oh, I'm just going to write a sequel. " It's not as simple. Sometimes you have to write code to insert a document. So, it's kind of a messy thing to do. So, yeah, so I want to use natural language. Let's head over here and then let's go ahead and do it. Um Add five treats for all users. And I'm going to send it in. So, what happens now is this data agent kit, not only it has access to the MCP server, it also comes with skills. The skill has the know-how, so it doesn't just communicate to my server, it also knows how to do that properly. So, as you can see right here, the agent actually suggests it actually wrote a Python file and it wants to execute this Python file. And it doesn't just execute because antigravity is aware of what's going on. It doesn't just go want to go ahead and just run application. It's not so secure. Yeah. So, it's going to ask for my permission. — That's good. And I'm going to go ahead and run it. So, it's going to answer five treats to three users. Awesome. We only have three users. Huh, I thought we had more. I don't know. They crashed on three users, huh? Oh, that's right. Oops. All right, so let's head over to the Dino Quest and see if we have the treat added. Awesome. As you can see, we have added five treats to our user. And as for the announcement, because we want all of you to know that announcements here and you know, with the

announcement, it's even worse because I also forgot where I stored it, which is not a good idea to code at 3:00 a. m. in the morning. Um so, you know what I'm going to do? I'm going to also ask Antigravity to do it for me again. So, here I am. I'm going to head over here and say add an announcement with title sorry for the crash with a message that says they each get five treats. And I'm going to ask it to do it for me. So, what this agent does, it's going to go through all my data store and luckily, this is not a very complex use case, so it was able to figure out where I'm storing my announcement. It will then create the document for me and even formulate a message for me. As you know, my English is not that good, so it's going to formulate a very nice sorry message so everybody can see. So, for those of you wants to see the message, if you go over here and refresh the message, there you go. The message has been inserted into the document. Isn't that great, Richard? That is amazing. You've redeemed yourself. Uh so, that was awesome. I mean, it's you're interacting with the database, doing things that not all of us have every skill, but you actually are baking skills into your workflow where you can do all these things. That's amazing. Those agent first workflows, though, if you see that, this is a mindset shift, right? This isn't how we're all used to working. There's something new here. And part of this, too, is thinking about this new way to debug. We're connecting to live environments, pulling logs, iterating on this. So, how do we think about this new approach and plan and and specifications, things like that. It's pretty exciting. But, let's look at kind of this next piece. Now that we have happy users, hopefully, you can try it again and keep messing around with it. How do we grow the app? How do we fix it, optimize it, make it even better, add performance improvements? So, our very second workflow we want to do is how do we optimize? How do we understand user behavior? How do we fine-tune the game to make it even more competitive for players? And the log data might actually tell us what kind of dinosaurs people are making, playing, whatever. But, instead of writing all the pipelines with ETL and analyzing data, there's probably a better way to do this with an agent. There's always a better way. Why do we want to do things manually, right? Like, that's just time-consuming. And so, by doing that, it's like we can use the data agent kit again. And here, data agent kit is an extension for our IDE, and we can see that we can have access to all these big query things. And so, in addition to using the data agent kit, we're going to be using the big query MCP, as well as a custom agent skill that we built for this actual purpose. And here, I have our skill open, just so it's not magic. You can see that we actually have a skill here, and it's going to lay out what it's going to do. This skill provides us a blueprint to automate a zero ETL analytics pipeline. So, what it's going to do is it's going to create a log writer sync to be able to stream the live logs directly into the big query data set. And that means no manual data plumbing that you needed to do. Then it's going to use the big query MCP to be able to run a series of parallel queries, and then just extract all that data that we got from the logs. And then lastly, it takes all of that data and it's going to create us an interactive dashboard. And so, question is how do we run the skill? Well, we can run the skill directly in Anagog Gravity. And as you see the theme, we can do everything all in one place. But because this takes a little bit of time because we have a lot of data there, I've run this beforehand cuz I'm pretty sure right now at what is it? 3:30 something, you don't want to be sitting here watching me run this. So, I've run this beforehand and all I had to do is tell it to run that report for my DynoQuest app and then it told me the actual pipeline that I had shown earlier. And then it goes and says it created a log report and it created our feature insights so that we know what we want from our next version and created an interactive dashboard. So, let's take a look at the dashboard that we have and here it is. So, this was what was generated. So, now we can see what the traffic overview, how many requests there were, um how many server logs and the win rate by dyno type. So, I probably should have cho- chose speedy earlier because apparently speedy is a 54% win. I know. I did choose tank. Well, at least it wasn't too bad. I balanced was the worst. And then how many treats per outcome and the top reused dinosaur. We could assume that if somebody uses this dinosaur a lot, it's probably a good dinosaur. Um but let's go back over here. So, in addition to building this dashboard, you actually now have all of the data from BigQuery. And you can just go straight into Antigravity and start asking it questions in natural language. No more need of getting sick, how do I build my SQL query? I mean, it's been a long time since I've actually thought about all the select statements and you know, all

of that stuff. And go and ask it. But in addition to asking it information, you can draw the distribution graphs just straight here. So, what's really nice about this is that you don't have to go click back to the console, do SQL queries, do, you know, come and give my graph and then you can do everything all in one place, no context switching. And I'm pretty sure there's no way that even though I said it takes a little bit of time of 9 minutes, there's no way you could have done this manually in that short period of time. So, what do you think about that? It's impressive. I mean, be honest with me, how much of a data person are you? Like, was a lot of this just helping you do things with the skills and the tools where this isn't your wheelhouse? Or how should people here feel like you're doing some pretty sophisticated things here? It's been a while since I've been messing with SQL, and you know, we do a lot of coding more for the front end, but not so much of all of this. So, for me of thinking about how am I going to put these queries is just something I just don't think about from my everyday basis. Amazing. So, we went from millions of logs and things like that or potentially into some sort of clear actionable features in a couple of minutes. That's the sort of leverage you hope to get from skills and tools in this agentic workflow. If I were building an app a year ago, I mean, I might be intimidated. So, I was asking you, it's a big query and doing these sophisticated queries, ETL pipelines, that's not my wheelhouse. But you're showing how using these skills and MCPs and tools, a lot of us can just get different work done. Which is pretty wild stuff. And that data agent kit, should check that out. It's all managed. All those skills and MCPs get updated. You get the latest and greatest whenever we improve that. All right. So, we've debugged. We've optimized with these agent-first workflows. I think let's go up a notch for that final demo. How do we think about an agent-first remediation, actually detect and fix problems? Yeah, I mean, nobody wants to get We all know crashes happen at any time, right? Like, right now there could be a crash at somebody's app right now. I hope not, but I'm pretty sure there is. And nobody wants to just drop everything that they're doing or if it's in the middle of the night, you know, Christina might be up at 3:00 a. m., but not everybody's going to be up at 3:00 a. m. to debug these issues. Nobody wants to be debugging. So, what if we did all the stuff manually? And I'm By manually, now I'm talking about us anti-gravity and doing it. It was faster, but we don't want to do that. — You're still involved. — Right? We still have to go do it. So, what if we did all of that 24/7 autonomously without us even watching? So, we decided that we are going to build a self-healing system. So, introduce to you our Dino agents. These are our custom agents. And of course, they're called Dino agents cuz we have a Dino game. Um you know, we It's got to be Go with the theme. And so, these are our Dino agents. There's the remediation agent, the CI agent, and the CD agent. These are three custom agents that we built, and they all specialize in different tasks. We hosted these agents on Cloud Run because this is where our Dino Quest game is hosted as well. So, by having it on the same platform as our actual app, we're able to get a secure, low latency, and cost-effective architecture. And this shifts all our manual troubleshooting into an intelligent and scalable assistant. These Dinos, these agents of ours, use uh the Agent Development Kit ADK as our framework. And you By using ADK, it saves us from writing a custom boilerplate for state management, tool routing, and retry logic. We also take advantage of ADK's skill tool set to be able to get the Dinos' expertise cuz we're able to load custom domain playbooks in order to codify all of that knowledge that they have into a repeatable work workflow. And what's nice is that all of these agents are production-ready observable agents. — [snorts] — So, we're going to see all our agents in action later, but let's focus on our first agent, the remediation agent. Our remediation agent is an always-on SRE and is able to catch and fix failures on its own and fix the root cause. So, the way that it works is that our app is always going to be logging into our cloud logging, and that's by default, right? Whenever it sees an error, it's going to trigger our event arc and our pub/sub. That is going to wake up our remediation agent, and our remediation agent is going to go and investigate the issue, and it's going to go fix the issue using Gemini. So, after all of that is done, that it remediation agent does its job, and it's going to go and open a GitHub PR with the fix, and it notifies you with Slack. So, all you have to do is, oh, I got a Slack notification. You don't actually have to be at your computer. You could actually have it on your phone with your

Slack. So, all of this is an asynchronous event-driven process that turns every error into an actionable job. Awesome. — want to show them how it works? — Exactly. That's where the agents work. So, while you two are, you know, uh talking right here, I uh deployed our Dino agents into the system, right? So, sorry, Richard. I have to break your app again. — Again? Again. Sorry. Just got back in my good graces. — breaking things. I have to do it. So, let's head over here and start breaking the system, right? But this time, I think we'll be fine cuz we have the Dino agent, right? So, if we come over here and look in here, this is By the way, there is a little code right on the left-hand side where you can see our Dino agents are working, right? So, let's see. Looks like our system is working hard trying to uh It's working hard to crash. — Crash. Right. Right. It's working really hard to crash. Your 3:00 a. m. system looks like a fever dream. I know, right? — That was pretty intense-looking. Ooh. Looks like it didn't crash for some reason. No. No? — You can't break it on purpose. I don't know I if I can. Maybe it's trying to self-remediate. — Yeah, maybe it already helped and self-remediated while you had to deploy the agent. — that's okay. So, we'll see how that goes. So, Oh. Did the Dino Quest Okay, perfect. It died. — It died. All right, sorry. Thank you. So, what you're seeing right here, feel free to watch this live. It's right here. There is This is how the agent works. So, you have the remediation agent listening to all the problems, and I'm going to just visualize it for you right here. So, the remediation kind of comes up and work, and it figures out, you know, your system kind of just crashed, so I'm going to help you remediate it. So, that's why it went off and remediated. But, not only does that, Yeah. remember what Denise was talking about like it went off and fixing the code, right? So, it went off and fixed the code. Not only fixes the code, it actually pushes things into our GitHub repository, right? — Awesome. So, as you can see here, um So, I don't know cuz our agent right here, you can see now, "Hey, That's right. — it kind of breaks, right? " So, you can kind of see it creates a leaderboard PR, so it tells us exactly what went wrong, right? It's cool. And it tells us, you know, "Since something went wrong, I'm going to help you fix it. " So, it even creates a very nice branch right here that tells you in this particular instance, this is the code that was fixed for you, right? So, very nice. And in the Slack message, I'm also getting a message coming from the Dino Quest, which, you know, kind of tells us exactly what went wrong. Perfect. Awesome. So, I hope you are following us at this moment. So, see this agents are working. Did you see when you were looking at the monitor, did you see that the agents starting to dance with each other? Yeah. That's the part where we have the A2A. So we are not just having one single agent, we have multiple agents. They're trying to work together. There you go. You see that? They're working together. Okay. Is that what that is? That's yeah, so that's why I have the swarms of agents working together. — See, they love working with you. — Settle down, agents. Exactly. Perfect. Yeah, so that's it. Um Richard, so you know what we This is what we have so far. Okay. As you can see, we have our agent. This is how it's built. So all these dancing funny agents and all that kind of stuff was built with ADK. Yep. And also it has everything that you need in order to make it secure. So every agent inherits I am um securities. So my CI agent doesn't go off and deploy things. Mhm. And my really agent doesn't have access to build my image. So everything is inherited, so it doesn't go off and do crazy stuff. And also because we are accessing a lot of external systems like the messaging system that we have back there and also the GitHubs, we want to store all the secrets somewhere very secure. So this is why we're storing it in the secret managers. Awesome. Isn't that amazing? That was awesome. Yeah. Cool. Woo. And because of Thank you very much. I know you're impressed, but you know, it's okay. Keep monitoring the agents. Um so if you all the agents are deployed on Cloud Run. Right? Yeah. And the Cloud Run, the good thing about Cloud Run is it goes dormant when it's when we're not using it. And when there's errors, it wakes up. And after it wakes up, what does it do? It fixes bugs for me. — Work. Yeah, exactly. Puts them into work, working hard for me, so I don't have to be you know, called in the middle of my date or my dinner with my friends. — Um and then when it's done, it will then shut down and you don't have to pay for it. Isn't that great? — That's the best.

So, I know money is great. I know all of you care about money much, but I'm pretty sure a nerd like you and hopefully me and hopefully you, really like to see how agents are built, right? So, this is one of the tools I want to introduce to you that how I built my agent. I built everything from ADK and building from scratch is never easy, right? So, you want to have some kind of tools to help you to how to build this. And I'm using something called agent CLI. The agent CLI It's Okay, let me put it on the top so everybody can see it. This is typing life, right? Make it a tad bigger for everybody. Can you zoom in a tad? There you go. Thank you. All right. So, this agent CLI allows us to It helps us to build, govern, also scale and optimize our agent. So, I'm going to So, do you want to see how I built the agent, right? Cool. Awesome. But, we don't want to be here all day. So, I'm going to show you how we get started. — Unless you want to stay here with me all day. — us off. Yeah, I'm sure. So, let's create a new agent. And what do you want to call this agent? Well, since we have dinosaurs and everything's dinosaur themed, let's do a T-Rex agent. You like the T-Rex name? Yeah. You don't want to spell Rex as Rex. — to think about it. What the T-Rex agent's actually going to do though? We'll figure it out. We'll vibe it. Anyway, so call it agent. Since I feel like I've been talking for a long time, so let's just go ahead and create prototype. Mhm. And then create. Yes. So, what it will do is going to create a boiler plate code that has everything there for me. So, if you head over to T-Rex and then let's see what's inside the T-Rex. It actually creates a lot of the things for me already. So it creates um an agent code right here. So if you can kind of quickly see what's in the agent. So you can see it creates an 88 code. It shows you what is in agent. It shows you how to build an agent with the instructions. So this is how your agent should behave and also a lot of the tools that you can include in your agent. So it's all there for me. And as well as it includes all the test. So I can use it to test and evaluate my agents. — That's great. Yeah, isn't it great? And also we also have all the um the files needed to containerize it so I can deploy later on. That's great, right? Amazing. Awesome. So now we have all the agents ready. Let's go ahead and hand it over back to you, Richard. Yeah, so agent CLI, check that out. Free, easy to use. You've seen data toolkit, a lot of these things easy to use, tap into. So more than that, I want to highlight this. This is something you all have been asking for a long time. I'm sorry it's taken a while. But how do we make sure we have more control over our cloud cost spend? So coming very soon are these spend caps. These are in private preview, but these help you set and enforce budgets on the APIs for things like Cloud Run, Gemini APIs and more. And what happens here is these particular caps pause traffic once your budget is reached. So you can't blow it up and then wake up in the morning and destroy your budget. So we're going to add this to more and more services, give you more confidence that you're not going to destroy your spending plans. Pretty good stuff. All right, so this was some great engineering. Agent CLI is pretty awesome as well. It's opinionated. Like you saw it had a set structure for e-vals and certain codes. So all those sorts of things you can still customize that as you want. But now that we've stabilized this game, is it completely bonkers to maybe ask for another feature, another level or something like that, we can do that, right? So, can we ship this code on another branch and ship a new feature of the app? We definitely — for you to do it cuz this keeps breaking, but we could try. Why do you keep on asking her for to do extra things? — keep testing her. Raise the bar. — So, yes, we definitely can. Um so, we created our own autonomous CI CD pipeline. You kind of already seen a lot of the details earlier, but before we're going to get into the details, you know, CI CD pipelines take a little bit of time. So, we're going to kick it off first and then we're going to talk about it. So, this time I'm actually going to kick it off on Slack, which means that I can kick it off from any at any time, anywhere. So, here I'm going to go in and do CI agent run CI on branch level two. So, going to kick it off. Oops. Let's go show that at the bottom. So, going to go on and start running this. But, so what we did is that we are while this is all running, we're going to look at the details and what we did was we have our last two agents that we showed earlier. We have our CI agent to be able to do our continuous integration. Um you want

to go next slide? Yep. Um and so, we have our CI agent with our continuous integration and then that agent is supposed to remediate all of the issues that it has as well as doing um a security scan as well as testing. Then we have our CD agent, which is our continuous delivery agent, and it deploys the app, it learns the deployment patterns, and it does monitoring with auto rollback. So, if there's any failures with the deployment, it's going to roll back to make sure that our app is in a good place. So, all of these agents talks to each other with agent to agent, A to A, and this almost makes it where it's like a team member collaborating. Rather than one agent telling some general agent some message to tell the other agent, it can actually talk to each other like normal human beings. We can also communicate with Slack as you see. We were able to get notifications. We're able to kick it off from the pipeline from Slack and we're using the Slack webhook that's available. And this, unlike static pipelines, this these agents are very intelligent and it can actually make decisions for us so we don't always have to tell it what to do. Cool. You want to show Of course. Yeah. These agents with the two agents that we built so far, right? And there could be more agents in your environment. And if you want to put them to work together, you need to know where they are. Yeah. And this is where why where we have the agent registry. And this is where we have looks like we our agent doesn't have a lot of friends. I mean, we'll give them some social work maybe sometimes skills later, but these agents we have three of them. CI agents, CD agents, and our cycle agents as well. So, if you look into it, so registry, so these are the friends and that tells you how to locate your agents. Where do they live? How do you call them? And it also provides a personality like knowing what this agent does, right? So, there's description of what this agent actually thinks and how it process things. Not only that, it also provides you with all the tools it has access to. So, now if I want to use this agent, I know what other things I can do with this agent. So, it gives you a much more clarity onto things that you use and to do that. Awesome. Perfect. Um right here. So, awesome. So, if we look at our CI agent, right? So, our CI agent is right here. And what it does, it actually takes in um notifications and it takes in um things to do from my other agents. So, it communicates with A2A, right? And it also take messages coming from Slack. And it also uses a lot of things together, right? So, it takes in decisions, it uses a skill to think about things, and then what it does, it use this information that's given from the Slack messages, from the other agents, it will determine if I want to run what kind of test. So, it will if you have a lot of front-end code, it will do run front-end test. If it does, you know, quick change some things, it will go through security, right? So, that's what it do. So, it's smart. It doesn't do a traditional pipeline, which is kind of does one and two and three. It's not sequential, because it thinks for itself. And then it's going to use a lot of the Google very well battle-tested services, like Workload Build, that helps you to build images, and also Artifact Registry to help you hold the images as well, right? And after it's done, it's going to send the um the notification or message to the CD agents, and the CD agent is our release manager. Uh when I was kind of coding this, when I was having fun with like running all these three agents, kind of they're running together, actually there's once where I made a small mistake, and the message well, I accidentally deleted one of the revision, and the CI agent was telling the CD agents like, um you know, can you please deploy this image that doesn't even exist? You know what happened? The CD agent actually went back and tell CI agents like, "Hey dude, what do you mean? I haven't seen this message. " And they um back and forth, until they figure out what's wrong, and the CI agent we ran the whole thing, and the CD agent eventually does the whole release. So, it's very smart. I was very surprised, and you can see them kind of doing the dance back and forth, cuz I was testing the image. It was super fun. Wow. So, the CD agents gets it um sees the changes, looking into all the changes in your code, and get decide on the risk score. The risk score is on how much What's the risk of, you know, deploying that into the environment and things can go wrong. So, it will use a score to actually determine the canary release. And it will monitor this for a duration, and since this is a demo, we'll look for like 30 seconds. Um and then we'll deploy everything into production. So, that will be That's the whole process of um

of our C of our CD pipelines. Wow. That's pretty cool system. It's learning. It's scary to me, but uh your system learns. Yeah. That's great. So, we showed that live, right? Throughout this session we've seen examples of Agentic tools doing different steps we see here. We plug this into antigravity thanks to MCPs, thanks to skills pulling in live log information, doing database queries. That's amazing. Now, you can build your own Agentic workflows. We showed you with that Agentic CLI, that Agent CLI. You can build your own things here, and I don't know. We won't want to always spend all of our time on day two. So, if we can use these tools to use the platform that backs our agents, it just makes life a lot easier. So, all the problem-solving that we love as engineers hasn't gone away. It's just shifted around. We're doing design work. We're planning work. We're interacting with the systems. I think that's pretty amazing. So, enough talking. Level two is shipped, right? You shipped it? It's there. All right. So, if you scan this time, we can see who can play now in the apparently terrifying volcano for these distressed dinosaurs, and see how this works. You have to pass level one in order to get to level two, but I got past level one. Yes. You have to pass level one. Getting through level one is hard. Shortcuts? Yeah. Left right left right? No. B A start? No. We should have coded that in. Awesome. So, you're going to be playing that in your free time. But look, if you want to incorporate these three agentic workflows in your own applications, you can scan these QR codes and actually follow the tutorials, which has all the code that we demo today. How awesome is that? They took everything they built and put this into a lab that you two all of you can run yourselves. So, I'll keep that up for just a second. Cuz again, this shouldn't just be something we all watched here. It should be something let's go try it. This whole event is about getting hands-on, building the confidence to go do some of this ourselves. So, take advantage of those code labs and see what you can do. So, let's go build. Let's go have some fun and uh thank you for joining us today and staying cool. —