Next gen agentic architecture: Hands on with Gemini 3.5 & ADK

Added: 2026-05-23

[00:00:00]Hey, welcome back. I'm Dave Elliott. I am a developer relations manager for Cloud AI here at Google. We're at IO 2026.

[00:00:09]Long day of live streaming. Today, we're very fortunate to be able to talk about some of the latest models that have been released. Gemini 3.5 and Omni. And with me today, we have Go ahead and introduce yourselves. Oh, thank you so much, Dave.

[00:00:22]Hi, everybody. I am Katie Winn. I am a developer relations engineer here at Google Cloud AI, specifically for generative media. So, anything that outputs, you know, images, video, music, text generation. That's the fun stuff.

[00:00:35]That's the fun stuff. That's the fun stuff. The media.

[00:00:38]Hey, everyone. I'm Lovie. I'm part of the same team. I'm developer advocate for anything agents, basically. So, today I'm going to show you some of the cool things other than the fun things.

[00:00:47]So, yeah. All right, great. So, let's dive right into it. I think we have about 20 minutes on this. So, next slide. Go ahead. Yes. You're driving. I am. All right, so um What did we announce here at IO today?

[00:00:59]We announced a whole bunch of things.

[00:01:01]These are the five things that I think matter the most. It's anti-gravity 2.0, Spark, and manage agents. We are not going to talk about those in our session. Yes. We get to talk about Gemini 3.5 and Omni. The fun stuff. The fun stuff.

[00:01:15]>> Yes. Yeah. And the cool things. And yeah. And we're going to see demos, hopefully. Cross fingers that work. All right, so go ahead go ahead and move on.

[00:01:22]Let's place some context. Let's create some context here. So, this is my favorite chart. It's kind of an eye chart, but it's what I call the chocolate bar diagram. It is everything that you need to know about building on the on the Google agent platform, Google Cloud agent platform. So, Gemini enterprise agent platform, we call it.

[00:01:38]Build, scale, govern, and optimize. Those are the big three pillars. Now, most people, I think, focus in on the build piece. That's that's part that really gets a lot of attention. But, I think the scale, govern, and optimize is is going to be, as we go into the future, even more important. So, let's talk about the things that we highlighted in red for the audience. So, Katie, can you talk about the Gemini models, the the Google models?

[00:02:05]Absolutely. So, since I do the fun stuff, I live a lot in the build section of this diagram. But again, all the other stuff is super important, especially as time goes on. But, accessing these models on Gemini Enterprise Agent Platform is, you know, super easy since I do the fun stuff, the media. I like to really do this within the UI specifically. So, you can go you can generate all the content, work within the UI, the Agent Platform Studio. And then also, developers can access all of these models programmatically via the API. So, you can do this through building agents with ADK, which Lavi will talk about, or through some awesome things like the Google Gemini SDK. Right. Right.

[00:02:42]>> Yeah. Okay, great. So, that's one little box, one little sliver on the chart. So, Lavi, walk through the the not all of them, but just the ones in red that you're going to show in the demo.

[00:02:51]>> So, so one of the very great things about this diagram is it tells you the real world that you go from zero to one with the build, and then when you want to go from one to end, this is where the scale, govern, and optimize comes into picture. And I'm going to show literally exactly the same journey today. But, the most important part is we're going to start with Agent Development Kit, which is the Agent Framework that we have, and you can build some of the coolest complex agents out of it. But, the other thing which I want to focus on is the Agent Runtime, which is very important because once you have the agent, you want this to be deployed somewhere so that your colleagues can use, your friends can use.

[00:03:26]And you also want that to be in a secure environment so that, you know, nobody hacks it. And this is where the scale part and the govern part comes in. The other thing which is the Agent Security and the Agent Registry because you may imagine yourself to be at a place where you have like hundreds of agents.

[00:03:41]>> Thousands. And thousands at some point.

[00:03:43]And this is where the registry sort of comes in very handy because you can sort of just get to that registry, figure out which ones are the most important agents that you have. And the last one, which is my favorite and and I'm going to show you a demo on this which is the evaluation because again, you can build agents, you can play with them, but unless they're evaluated and the organization sort of agrees and aligns with that, there's no point of doing that.

[00:04:05]And that's pretty much what >> Yeah, and things change. Like 6 months from now, the evaluation you did on day one may not be relevant because you have a new vendor that used a different language.

[00:04:14]>> Right. You you you really have to control the context and and you'll talk a little bit >> Yeah. And and this is this is just couple of things. As you can see in the diagram, there are so many other things that we don't even have time to talk or show. But yeah, that's the beauty of this diagram because with agent platform, you basically get access to all of this and and couple of things that I'll show in my demo that how easy it is for you to get started with that journey. So it's not that it looks complex and oh my god, you have to learn so many things. It's really easy and and I'll show that. All right, great. So let's move on to the meaty meaty part of this. So let's talk Gemini 3.5. So this is the big announcement that we had today. Gemini 3.5 we say achieves, I'll read it here, achieves higher intelligence while bringing down the cost of the frontier.

[00:04:58]So really what the I think the three things that that we probably sort of stake 3.5 on is improved coding.

[00:05:05]Yeah. Um agentic workflows or it's better in in and you'll talk a little bit about that what we mean by that. And then multimodal. Multimodal has always been a differentiator for the Gemini models, right? Trained from the very beginning to to be multimodal.

[00:05:19]So this is what we announced yesterday.

[00:05:23]And this chart on the upper right hand side here, this shows basically my interpretation of this is you you get near state-of-the-art performance, near state-of-the-art results. Let's say performance. I should should be specific. Results at much lower prices, lower cost. Now while technically the cost the price prices of flash have increased but relative to the outcome, the the the results, it's really giving you close to that state of the art at a much much reduced price. Those are the two big things.

[00:05:57]>> And and you know one of the things that Sundar mentioned yesterday in the keynote where if you actually do the calculations, you end up saving billions of dollars just by at that scale. So, while the cost may seem slightly higher compared to the previous one, but the amount of intelligence that you get at that price, I think that's really impressive. And the other thing which I'm going to I know that you're going to talk next is the token per second. So, Well, no, actually the next one is is exactly what you're saying. Go and go to the next slide, which is the benchmarks.

[00:06:22]And these these benchmarks are the total I charge. It's it's seems like it's obligatory that we talk about these every time we launch a new model. All the vendors talk about it. But a few things that actually will stand that do stand out that we're going to show here and on in the demos are the the coding, right? And the Agentic. And you talk a little bit about those how how much better they perform.

[00:06:42]>> Yeah, and again, just to put it into the context, this is we are comparing with the 3.1 Pro model that we had and the three flash that we used to have.

[00:06:51]From there, we have done a fabulous job with terminal bends. Which essentially means in a very simple way that now the models are trained so good that it can do a lot of terminal batch level things in a very good way. And you can see we have done a great improvement there.

[00:07:04]The Sweet Bench is again a industry standard and you can see that we have done a good job. Again, just to add a context that this is a flash model. This is not the Pro model. We still have to announce that and at some point you will see that as well. But a flash model with so much power, you you get the industry standard Sweet Bench Pro results. And then the the MCP Atlas, which is again the way it sort of calls MCPs and how efficient it is. And you can see the numbers how we have gone from 78 to 83.6. This is really good. And the last one, the tool one, which is again great for 56%. All of this means and again, there are other numbers as well, but the most important thing to remember and take away is that this makes your Agentic experience with any coding harness, whether it's anti-gravity, anti-gravity CLI, or any other that you use, it's going to you're going to really feel those differences in those areas. Right, right. And you know, the last piece, since the name of the model is Flash, is the next slide shows really the performance. And this is uh uh you know, layered on top of what I said performance, um layered on top of the outcome, right?

[00:08:07]The the better results, um the more uh usable results, is how quickly they get they they get um returned. They're done.

[00:08:15]I think that really is one of the things that stand out from this model.

[00:08:18]>> Goes goes with the name, like this is this is I think if you see the bottom chart, and this is my favorite chart because you can clearly see that not only do we match up the performance to the 3.1 Pro that we used to have, which a lot of our customers and developers use, we are so fast now, and again, I'm going to show that um uh that how fast this is the moment you sort of just put a query, and uh it's it's really fast.

[00:08:40]And and that's the kind of thing that we're aiming with the Flash models, which is you get the best performance in the cost which sort of is slightly cheaper compared to others, and you get very fast open per second. Yeah, exactly. Um let's jump We've teased at your demo a lot, so let's let's switch over now to Omni. So, Omni was announced yesterday, um and Katie, would you like to give us the snapshot on what Omni is and why it matters? Yes, so Gemini Omni, um it's really awesome. My realm of generative media, very fun. So, right now it's a new model in a Gemini family of models called Gemini Omni, which the idea is you can see on the screen, you can create anything from any input, but starting with video, so you can take any kind of modalities and translate those into videos.

[00:09:22]Awesome, awesome. So, you have some demo-ish things to show us? I do. I do.

[00:09:28]I've been playing around with this model a lot. Um you know, users can go out and try it right now on consumer platforms, so Flow, the Gemini app, or YouTube Shorts, and it's coming soon. We announced in the keynote yesterday to APIs and developers and on Agent Platform.

[00:09:44]But, I've been playing around with a little bit, so I kind of wanted to show you some of my creations. All right, let's see it. Um yeah, so one of my favorite things about this new Gemini Omni model is it has all of the world knowledge, you know, think about all the benefits you got from Nano Banana when that, you know, became a the architecture that it is now, and you have all the world knowledge benefit of being a Gemini model.

[00:10:03]Uh so, [music] I took this suitcase, for example, and I gave it this prompt and I said, you know, float up and spin in circles, which is, you know, unrealistic. Suitcases can't do that.

[00:10:11]But, then it like opens up in a realistic way, and it has the physics and the knowledge behind that and can render different text within that. So, in this case, I called the suitcase the Omni case. Yeah. Uh Clever. Clever, right? I know. I I didn't even use Gemini to come up with that one.

[00:10:25]>> [laughter] >> And uh add some fun colors and stickers.

[00:10:28]So, so you can see that video here.

[00:10:35]It's really good it it remembers the exact context of what you've actually given, which is very interesting.

[00:10:39][music] Yeah. Yeah, so the beauty of Omni is, you know, it has all of the same modalities that you're used to with a video model for video generation. So, this case is an example of, you know, starting from an image and using that as the first frame for your video. So, if I play it really quickly again, you can see it starts in that same [music] first frame.

[00:10:57]>> Yeah. Uh spins it around, opens it up, and gives some fun text and And that took you that took you all of a but 5 minutes?

[00:11:08]Yeah, from from idea to conception, definitely. The the speed on this model is is really awesome, too. We're talking about >> but the fascinating thing is you you actually wrote Omni case with fun burst of colors and stickers that pop, and it actually did that, and it was very fun like very interesting that it was able to remember and Yeah. And it's, you know, this isn't the best grammar by any means in these prompts. You can tell Gemini didn't write these. Like, they came from me. But, [laughter] um you know, it's not super grammatically correct, but this model has a really deep understanding of text prompting and you know, trying to get out what I want and it you know, every time I ran this prompt it renders the Omni case exactly correctly in that same type of logo that I was after.

[00:11:45]>> I just a funny funny story for for for the audience. So we're here at IO and we talk about world understanding and there's a photo booth and an photo booth and I went in there with this shirt on and one of the options was show show me presenting at IO and I thought this would be fun and it sort of put me on a stage presenting and what I loved about it was it showed me presenting a session on ADK with this t-shirt with that logo up there. So it knew that this what what this logo was and that's something that is reasonable for me to be showing. So that that world understanding now comes to video. Yeah, absolutely. And that's a great example, too. Like it just had the logo. It doesn't even the model didn't even see ADK on it on the shirt. So it's able to to have that context.

[00:12:28]>> I can only almost imagine the amount of branding and marketing that people can do with this. Like you can now I don't think so it was possible before in the models where you can sort of maintain the context, maintain the structure of what you're trying to >> good that's a good segue to one of your demos.

[00:12:44]>> of my next demos. We're going to take a slight detour along the way cuz these are the order the slides are in. Haha.

[00:12:50]But the other great thing about about Omni is you know, the world you know, we talked about all the context. So if you say things like you know, create a claymation video about Newton's first law of motion and I don't need to explain to the model what that is or provide any kind of dialogue and this is the output from this prompt.

[00:13:06]An object at rest will stay at rest and an object in motion will keep moving forever.

[00:13:15]Unless stopped by a wall. That's so cute. I wish we had that when we were studying. Right? I know.

[00:13:21]>> so many videos.

[00:13:22]>> I know. School might have made a lot more sense at the time, but >> [laughter] >> And the great thing about Omni too is it has outputs of 10 seconds. So you're able to have a little bit longer videos when playing with video generation as well.

[00:13:34]And then you know, going back to the whole theme of Omni, you can take any modality in. So you can take this video that you just generated and regenerate it in different styles. So maybe I want the same concept.

[00:13:58]Very nice. Yeah. And getting back to Lobby's point, you know, we talked a lot about what could this could mean for marketing and personalization in different scenarios. So you know, I said, you know, what about me? You know, what if I put myself in in a video? So What could go wrong? [snorts] >> What could go wrong? I know. So I took me with my dog Remy that is more well known at Google than I am at this point. [laughter] And he's nodding his head.

[00:14:22]A nano banana storyboard of kind of, you know, I asked Nano Banana to imagine Remy and I having a day at the beach. So hot here. I kind of wish I was by an ocean. So this was as close as I could get, but this is the the result.

[00:14:42]This is really cool.

[00:14:46]Like your face doesn't change. Like it's so consistent with your face. Yeah.

[00:14:50]>> That's really cool. No, the consistency and kind of what you were saying earlier is like a huge unlock with these models.

[00:14:55]You're able to provide kind of anything as reference. I have, you know, different character images, but I also have an image that's just a whole storyboard. Even with text in the image from Nano Banana.

[00:15:04]>> We talked about this at lunch where you like the the video generation quality is so high now. But the real pain that people have I think is in the prompting and the editing. And this sounds like feels like it's like the nanobananafication of video.

[00:15:20]>> Yeah. All of the video.

[00:15:21]>> Yeah. In in fact, I think the storyboard idea is very fancy as well because you Yeah.

[00:15:40]Yeah, you can have a lot more control over planning out the scenes and you know, switching between them in the way that you want them to be presented just through images through if you have like a video.

[00:15:49]>> Yeah. You can supply that as an input.

[00:15:50]You can do audio as reference and so there's a lot of really creative ways that people have.

[00:15:54]>> Would you want to say anything to anything? It's exciting.

[00:15:56]>> Yeah. It's exciting. Now, we only have about 9 minutes left. So, I want to make sure we have time Yes. to close this out and >> [laughter] >> He talked a lot about it so we should we should see it.

[00:16:06]>> Never want to follow generative video though because that's pretty cool.

[00:16:09]>> Okay, and I'll leave you with with one more Lobby just to make everyone jealous here. So, I'm obsessed with my dog. We all know that. So, here's a video of him um on a G bike going into a Google office.

[00:16:21]And kind of to Dave's point, uh real unlock of this model is you're able to do iterative video editing through natural language. So, I said, you know, that was that's the prompt I gave. Everyone can read it and this was the result.

[00:16:36]>> [laughter] >> Oh, it has the Noodler hat. Yeah, I had a a hat and so it's you know, really consistent in the video as input just with with different characters. So, Outstanding.

[00:16:46]>> Yeah. I can't wait to play with it. I know. I know. Can't wait to see what everyone creates. All right, but now let's go back in the remaining time that we have and let's talk about Gemini 3.5.

[00:16:54]So, again, the differentiators were around agentic, were around coding and then as always multimodal. So, I think you're going to show some of the agentic and coding things, right?

[00:17:02]>> Right.

[00:17:03]And just to set the context, so the choice of agent development that we have across the Google ecosystem as well as agent platform are these. The the one that I'm going to be specific to is the first three which is the agent development kit, agent CLI and anti-gravity. So, let's let's go to my screen and let's let's build something.

[00:17:21]Okay, so what I have here is a very blank setup of a project. As you can see on the left, I have bunch of the prompts that I want to try. I don't want to type just in case. And then I have a Gemini.md which has like some very basic tool basic tools and stuff.

[00:17:38]And I did this just so that you know, because I'm doing this live and as you can imagine that things can go wrong.

[00:17:43]So, I just wanted to add certain rules.

[00:17:45]Now, here's the thing and and this is where we started when we were initially discussing Dave which is the build part of the whole platform is very important because this takes you from, you know, having a very rough idea of what you want to do. Maybe there's an agent that we want to build with Omni and all of that. So, that's the first thing I want to show which is I have two things set up in the system. One which is the agency CLI. Agency CLI is a CLI that we offer which has access to bunch of different things that you can do with agent platform from an agent perspective and then it has skills for EDK. What it does is it basically knows how to build a best EDK agent. And then when you give your ideas and this is what if you see the prompt that I'm doing here which is build me a daily news bot using EDK. And this is a project with Dave's keeps asking you which is hey, can you tell me what is the news that we have today? So, I thought why not let's let's try to build it.

[00:18:32]So, what I'm doing here is that I'm I just asked this that here this is what I want. I want RSS feeds of latest AI news summarize the five stories. And the last part of this is very interesting. I say I should be able to deploy this and fetch the latest stories with what I want. So, I'm trying to define what I want. This is basically from zero to one. Let me hit enter and you can see that I'm using Gemini 3.5 flash the medium version because you're not trying to do anything complex. Now, while this happens, the one thing that I want you to notice here is how fast this is. And the first thing it did is it it figured out that I need to figure out the agency CLI and I need to know how to use it.

[00:19:05]And that's the part which we were discussing even in the even in the benchmark which is how good it is to do the tool calling, how good it is to sort of do the agentic, subagentic thing. And you can see that we are just sitting idle, we are not even doing anything and it just figured out what I need to do, how I need to use agency CLI. It called bunch of different sort of MCPs, tools and all of that and this was really fast. The best part of Antigravity ID that I like is that it comes up with the implementation plan. So, what you see right now is that it figured out everything and it gave me things like, oh, so I'm going to do this and this is the deployment permission that I need to do and this is the whole setup of what the final agent sort of looks like. Now, the best part I can do in this is I can actually add comments.

[00:19:52]You know, this is very similar to how we do things in Google which is when I am trying to build something, I write a doc and I say, hey Dave, do you think this is good? And Dave can comment that, oh, no, maybe do it other way or something.

[00:20:02]And that's the beauty of it. It's a It's a usual comment. So, you can actually add this and then what it does is it sort of keeps that in mind while doing this. So, I'm not going to change much of this. I'm just going to hit enter or proceed. And you can again see that it it basically created the task. So, if I click on this, you can see that there's a task list that it has created. And on the left, what you will start noticing is that it basically created the whole scaffolding for the agent. And this is interesting because you just started with an idea and then it figured out that, oh, I need the app folder, I need the test folder, I need bunch of these different files. And that's what I meant that now going from zero to one makes it really, really fast because of access to agency CLI and access to ADK skills and all of that. And again, I'm not doing I'm not touching anything. You can see how fast things are on the right.

[00:20:52]And this this really puts the whole 3.5 into the context because you can see A how fast this is, B sort of it knows which tools to call, what time when to call and how fast I need to sort of reference this.

[00:21:07]And that's pretty much what's happening.

[00:21:09]So you can see right now it's sort of trying to UV sync. So it's doing the installation.

[00:21:13]Now again, just for the context Dave, this is again something that we've seen so many developers struggle where we go out and we teach about ADK. And the very first 20 minutes this goes in installation, you know, talking about ADK and all of that. Imagine that we just say this, this will just unlock so much of what they can do with this.

[00:21:31]Yeah. Yeah. Speaking of a very practical way for developers to get started and just to to see some to get to some value. Yeah. Yeah. And and again, while this is sort of I want to quickly show you that if you see on the left, you do have bunch of folders. One particular, if you see the app folder, which is where the agent.py again, this is the ADK way of writing agents, which I assuming that let's say I don't know, it just created the whole agent.

[00:22:03]What you see right now is that it was able to figure out that hey, I did some mistakes in the agent. This is how fast it is. And it was it's now changing all of those things. So you can see how it's now doing the rough check. It's now doing the evals. And that's one of the point that we were talking about, which is evals are so important because the evals can change. Agency I with ADK skills sort of knows this. And whenever it builds something, it has a eval as a test. So it it knows how to keep testing and keep sort of improving it unless it sort of passes all the evals. Okay, so we we we have about 3 minutes left and we're going to be able to get to The final part? Okay.

[00:22:42]While this sort of goes on, and I think it's almost at the end of it. Oh yeah, I got the post.md. So finally, in the good time, you can see that I got the post.md. This is basically, if you see all the best sort of AI news across the internet. I didn't define what sources, it figured it out from that. And that's it. Yeah. So you were able to build all of this. Now the other two prompts which I didn't include and we didn't have time is you can actually now just say that hey, just deploy this and it's going to just do all the deployment for you.

[00:23:11]Of course, once the auth is done on behalf of you. So yeah, that's the beauty of 3.5 with ADK and the agent platform. So you raced through that demo. That was awesome.

[00:23:21]Would love to have seen the other two parts of it, but I think to summarize, you know, you showed us with 3.5 some of the potential 3.5. What I always tell people with any new model is just go try it.

[00:23:32]Just go play with it, experiment with it. You know, this is as as true today as it was, you know, three or four years ago when we were getting new models. We we we don't necessarily always know the potential, the possibilities. You know, we are going to learn as we Yeah. do and then iterate. Some things are going to be better than others.

[00:23:51]And so my advice is please go ahead and give it a try.

[00:23:55]Okay, so let's wrap it up here.

[00:23:58]Lavi, what's the best way for people to get started with agent platform and Gemini 3.5? So the best is you can just go to our docs.

[00:24:08]We have bunch of amazing docs around all of this. Agency CLI has a doc, ADK has a doc, agent platform has a doc. So just go there.

[00:24:16]And the the most fun part is you can just point all of these docs to agent anti-gravity and just let it do the things for you like I did. That's awesome. That's awesome. And of course you missed the opportunity to show shirt ADK.dev. ADK.dev. No, I didn't have a memo about the shirt. And [laughter] the best way to get started with with VEO today is you mentioned it before. Yes, you can use VEO and any of our other generative media models within agent platform.

[00:24:41]And Omnis in >> And Omnis in Flow, YouTube Shorts, and the Gemini app and will be coming soon to agent platform.