# Google's Agent Upgrade

## Метаданные

- **Канал:** Sam Witteveen
- **YouTube:** https://www.youtube.com/watch?v=TmqI-pX9aho
- **Источник:** https://ekstraktznaniy.ru/video/22374

## Транскрипт

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

Okay, so in this video I want to look at the new drop from Google Labs of updates to Opel. So remember Opel started out as like a drag and drop agent builder where you could sort of define different rails how things were going. But the latest update introduces a number of patterns that kind of show where agents are actually going. And it really shows as the models are getting better, the way that you actually build agents with frameworks and harnesses is dramatically changing as well. So let's have a look at what's new. I'll compare it to some of the things from OpenClaw that have become really popular and then we'll also have a play with it so you can actually see what you can actually do with this. So first up, just as a quick aside, the Google Labs team is like shipping like crazy. Google Labs is the team in Google which basically is tasked with coming up with new ideas and new products especially around things that use the Gemini models that use the generative AI models. So obviously their biggest hit to date is notebook LM [clears throat] but you can see just looking over the past couple of months they've shipped a whole new update to flow and how you actually sort of build movies and stuff like that. They've shipped updates to things like Pomelli for basically making use of the nano banana models and how to get the best results out of those. And on top of this, they've also shipped updates for Stitch and a whole bunch of different things that are more and more focused on giving sort of the average user the ability to create highquality outputs with the latest generative models. So, if you're looking to build any kind of generative AI app yourself, you definitely should be checking out what they're actually doing. Now, if we take a look at the latest Opal release, for those of you that don't remember, Opel is basically what Google calls their no code visual builder. Like I said before, this started out as sort of a drag and drop thing. It's now morphing into an agent's done for you kind of tool. And just recently, they also made it so that you could actually use the opals that you create in the Gemini app. So this does look to be one of Google's plays to look at how you can actually give normal people who are not involved with the cutting edge of agents access to be able to make their own agents do different things. And it makes sense. Google is basically trying to work out what are the key things that people actually want to use agents for. And if they can give people access to make their own and see what kind of agents become really popular, this is definitely then something that they can think about both when they're designing new products, but also when they're creating new versions of Gemini and of the other generative models. All right, so jumping in here, we can see that there's a number of sort of key things in this release. So they talk about that they're introducing this new agent step that turns static workflows into interactive experiences. And for me, this is a real clear sign that as the models are getting better, you don't need to keep them on a sort of single track path as much as you did before. So whereas in the past you basically had workflows that were step by step that were planned out and what I've often referred to as being agents on rails, the latest models with Gemini 3 and with things like the latest 4. 5 and 4. 6 six versions of Claude, the models are actually getting much better at being able to make the decisions themselves of the best path to reach a particular outcome. Now, a lot of that's got to do with things like planning and things like memory. And we've seen this already in things like clawed code. We've also seen it in anti-gravity. And even in things like OpenClaw. One of the reasons why OpenClaw became so popular is that people can just go in there and tell it what they want it to do and it starts building these sort of mini agent workflows in the back end which can be triggered either via sort of chron job which they often call a heartbeat or be triggered by the user actually asking for something to be done. Now you can kind of look at this as Opal actually upgrading to be able to do this kind of thing and you can see this clearly here in this idea of a generate step. So it talks about this agent step proactively determines the path based on your goal triggering right tools and models. Now I'm sure that this is not going to be as wild as open claw and there'll going to be a lot more sort of security and safety things in there that if you do give it access to data you're probably not going to have some of the issues we've seen with open claw where people have had their data wiped and experienced all those sorts of things. But don't be fooled. This is a fundamental change in the way that we're building agents. And in the past, we've seen this sort of play off in different frameworks. We've seen the early versions of Crew AI be totally off the rails and actually be really not that useful at that time with the levels of models there. And then we've seen the sort of counteractive movement of things

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

like Langraph which have been heavily on the rails where you can strain exactly what the agent can do. And I think we're seeing even with things like deep agents from Langraph and some of the more sort of modern ways and harnesses that people are using to build agents that there's a clear acknowledgement now that the models are getting good enough that you can let them be off the rails a lot more than you could before. So another key component of this release and of this sort of progression is the whole idea of memory. Now unfortunately we don't know exactly how Google is doing this. They're not sort of declaring this but they are telling us that your opals can now remember information across sessions and this makes your opals grow smarter and feel more personalized over time. So memory is definitely a thing that we're seeing with the latest sort of harnesses and different kinds of agent frameworks as becoming extremely important. The way that something like openclaw approaches this is mostly through markdown files and JSON files and that works really well because it's basically a single user agent system and we are definitely seeing this sort of change between single user agent systems versus multi-user agent systems. Another feature that they've added in here is the whole sort of idea of dynamic routing of where you're relying on the model more to decide how you're going to traverse the actual graph of nodes and stuff like that you're actually doing here. Now, this is more like something from sort of langraph, but adding it in sort of a more consumer-like product like this is actually giving people the access to make these agents where they've got a wide variety of choices. The last thing that they're putting in here is what they're calling interactive chat. And really what this is basically human in the loop. So we know that sort of like the best agents ones that realize when they get to a point where they suddenly realize that okay going this direction is not going to work or I need to get more feedback from the user. So here you can see this is described as sometimes an AI agent needs to ask follow-up questions. And now this agent step can actually do the human in the loop sort of thing. So while the sort of consumer version of this is being called interactive chat, don't be fooled. This is your human in the loop step of where you go back to the user, you get more information, you have some kind of chat with it. And it certainly makes the product much more reliable at the same time as giving users more options for this. So, I'll jump in and have a play with this in a second. But overall, what I think is really interesting here is that this is a consumer-based product that we're seeing from one of the big companies that is starting to use some of these key things that we've been talking about in building agents over the last year or so. And we're seeing that they're able to do this because they've now got better models that actually work with each of these particular steps. So, let's jump in and have a look at actually playing with it. Okay, so first up, you got a choice whether you can sort of remix a pre-made Opal or you can start from scratch. So here I'm going to start from scratch. I'm basically just asking it, can you create me an Opal that uses search and the web to find events, activities that are on in a particular city over the next week? So let's see how this actually goes. All right, so we can see that actually this built a lot quicker than I thought it would. So okay, it's what do we've got here? We've got our first node which is going to be the city name, right, that somebody inputs in. Then we've got a set of nodes where we're going to find city events and activities. And if we look on the side here, we can see the prompt for that node is here, right? And my guess is that this is going to have a set of tools that we can actually use. And sure enough, one of those is going to be search the web, get weather, get a web page, search maps. Quite a number of tools that we've got in there. Next up, we've got generate a comprehensive event list. So again, we've got the prompt. We've got our tools in there. And then finally, we've got this thing of render the event listings web page. So we've got a bunch of information in there about the layout, about the style design language, about component guidelines, etc. So let's try it out. All right. So I can just come over here and I'm going to type in Tokyo. All right. So I'm actually headed up to Tokyo next week. So, let's see how this goes at finding what's on over the next week in Tokyo. And we've got some nice animations going through what this is. Now, obviously, I could come in here and customize this if I wanted to make my own nodes and add in more things and stuff like that. If we look at the console, we can see that, okay, I put in the city name, it started finding different things. We can see what was sent to the model and stuff each time. And then we can see also what the actual steps from the agent coming back are. So we can see that okay, it got some structured data out by the looks of that. It's had some thinking to

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

itself. It's structuring those events. And now it's in the process of actually rendering all this out. Okay. So let's go back to the preview. And we can see sure enough, okay, it's given us a city guide. It's talked about what's on. So, okay, this is a doll festival. Probably not my thing, but we can see that there's a lot of things going on like that there. And at this point, you can see that maybe I don't really want things aimed at kids or farmers. So, I'm going to ask it to give me some more information based on the kind of activities that I want in here. So, I could come back and edit these steps. So, I'm going to edit the steps. We can see it's definitely got some interesting things like the early cherry blossom viewing and stuff like that. But we can see that I've asked it now as well as getting the city name, ask the user what kind of events they're interested in and if they are taking family, etc. Okay. Now, interestingly, that update took a lot longer than it did for the first time. So, it seems that a lot of people are actually starting to use this and they do mention up here that they're working to get increased quotas just for the models and stuff like that. So this is it is kind of amazing that Google is the company with probably the most amount of compute in the world and yet still they're struggling to get enough compute to run all these Gen AI kind of things going on. Anyway, we can see now that we've got different nodes for the start. We've got our city name, we've got the events interest, we've got a family status in there, and we've got the other things actually sort of updating to this. So, looks like we lost one of the nodes where it's been joined. Anyway, what we will do is we can try it now as an app. So, you've got the editor here where you basically try things out. Remember, you can manually come in and add different things in here as well. You can add in user inputs like that. I'm going to use that. You could add in a sort of LLM call or a generate and that can be a variety of different things. that could be calling Gemini 3 flash nano banana audio LM LIA not the latest LIA and I'm not even sure which version of VO that they're using in there but you really have the ability to make a lot of things in the editor then if we come to the app and we now try it and I say right I want to have Tokyo what kind of things okay this time I've basically put in art events music and food festivals and then we can see again it's going off and doing the thinking and doing the actual steps. And if we come back in here, we can actually see that console again of where the steps are. Looks like it's gone through that. It's now generating the HTML. And now, obviously, this example is not a complicated agent, right? It's very just a reasonably simple workflow. But with the new way of doing it, you can actually make a lot more complicated sort of agents where you've got these decision points where it can actually kind of decide different directions in the graph and go through them. All right, so you can see this time we've got probably a lot better kind of thing. Okay, the Tokyo Indie Music Festival. We can see now we've got a web page with all of this and it's been made specifically for us. And while this is a simple example, you can see that from this, nobody needed to code. Really, in many ways, nobody even needed to understand that whole drag and drop system. You could just create it with text. And I'm sure at some point you'll just be able to talk to it and make it happen that way. But you can then build this sort of little Opal app that you could then basically share it. I could then publish give it to other people. So, it's definitely worth coming in here and just checking this out and getting a sense of like, okay, what can I actually do with this? Now, they've got a bunch of ones that are already made. You can see, for example, here's a Google calendar opal. You can now even uh take ones like this where okay, it's basically taking a YouTube, extracting the transcript, analyzing it for educational content, and then generating a quiz and displaying the quiz in there. So, if you've got kids that are trying to learn or trying to do anything, this is definitely something you should be checking out. And don't forget, this is free. This is not costing you anything to try it out. You're not having to pay for the Gemini calls and stuff like that. If you've got a Google sort of Genai account, my guess is that you get more quot and things like that. But for trying these things out, you really want to come in and just have a play and sort of get a sense of what is possible and what you could just build as little Opals and little apps to basically share with your friends and family. So come in and check out what's been released with Opal this week. Get a sense of where these things are going cuz already we're seeing other companies make things like this for their own particular use case. And if you're someone who's actually looking to actually build agents in a corporate setting, this is something definitely you need to be aware of. Anyway, as always, let me know in the

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

comments what you think. Have you tried this out? What did you like about it? Tell me what you didn't like about it. Would you like to see something like an open-source version of this that could work with any models or other connectors, etc. Anyway, and as always, if you found the video useful, please click like and subscribe, and I will talk to you in the next video. Bye for now.