Developer Keynote (Google I/O '26) - Audio Described

On the stage in Mountain View, California, Josh Woodward strides out as an AI generated speaker name card shows an animated version of Josh jogging across a basketball court in an open book. >> All right, welcome back developers. How's everybody doing? >> All right, it's great to see you again, and it's wild to think back about where we were just a year ago. This morning, you've heard all about the advancements across Gemini. We've got the new Omni model, the new Gemini 3.5 series of models, and last month we also shipped Gemma 4 under an Apache 2 license. >> Josh points to someone cheering in the audience. >> It's our smartest open model yet. It's purposebuilt for advanced reasoning, agentic workflows, and the response has been incredible. 100 million downloads in the first month and it's pushing Gemma downloads past half a billion. So, Gemma 4 packs massive intelligence into a footprint small enough to run offline on your phone and it's deployed in everything from robots to satellites in outer space. And as you saw this morning, the big shift is our move towards agents. From AI that simply assists you to agents that help you get stuff done under your direction and faster. At the center of it all is Google Anti-gravity. It's our Agentic development platform that enables you to build with agents. The core platform is the anti-gravity agent itself. And today we're going to show you how you can build with it anywhere on our infrastructure or your own Android or the web. And when you step back, you can see how all of these pieces are coming together across Google. From the cutting edge models to the agent tools like anti-gravity to the platforms where users experience the magic and to the infrastructure that makes it all possible. We're working up and down the stack to help you build. So, let's jump into some demos and keep the launches rolling. Please welcome Logan to the stage. >> Logan Kilpatrick steps up to a desk as an AI generated speaker named Croch Logan leaning out of Whimo's window. >> Thanks, Josh. Our latest models help you build for the next era of agents. And we're putting the full power of agents directly into your hands, no matter what you're building. Need complex research done? Check. A data science task check your own personal radio show. We'll get to that one. Last December, we introduced the interactions API. It gives you one simple and powerful interface for both models and agents. Deep research was the first agent we made available through the interactions API. It showed what's possible when you give a model a goal and the freedom to go find the answer. But deep research was just the beginning. Inside Google, we're using the anti-gravity harness to empower our agents to intelligently tackle the most complex tasks. It's the same technology behind Gemini Spark and our coding agent in Google AI Studio. And today, we're making that same anti-gravity harness available to you with managed agents in the Gemini API. You can easily start creating your agents by adding custom instructions, tools, and data through a single call to the Gemini API. But building agents is one part of the puzzle. In production, these workloads need secure isolated execution and solving that on your own is an infrastructure nightmare. That's why every managed agent comes paired with a remote Linux environment hosted by Google. In a single API call, you get an agent and the sandbox. We handle the infrastructure so you can focus on building. We've been testing this with developers both internally and externally. To give you an example, the team at Stitch, our Vibe Design product and labs, used managed agents to enable their users to import their projects design system directly from their codebase. Their managed agent connects to a user's GitHub repo, analyzes the codebase, and generates a readytouse design. MD file, giving Stitch everything it needs to create onbrand designs. And because it's a managed agent, we handle the infrastructure, so Stitch can scale to millions of users. Now, Stitch is just one example of how powerful managed agents can be. Externally, early access customers like Ramp, Resemble AI, and Clippy are also seeing the power of building agents at a pace like never before. And we can't wait to see what you build. Managed agents are available starting today. And thanks to our ecosystem partners, you can start building with your preferred stack from day one. Let's jump into a demo. I'm here in the AI Studio playground where earlier this year we added support for interacting with agents. To help you get started, we've added a set of new custom agents that you can run instantly. >> Logan opens a window filled with AI agent templates. >> These open source agent templates come preloaded with instructions, skills, and tools as markdown so you can customize the source files like in this customer support agent. >> Logan chose the code then returns to the previous page. So, I told you I'd come back to my own personal radio show. So, let's select AI talk radio agent. This agent can produce a complete talk radio show about anything you want. You give it a topic and it handles the rest. Of course, I want mine to cover today's top news. So, let's see what the hacker news comments have to say. I've asked it to generate a fiveinut radio show based on the latest from hacker news. What's happening right now is the agent's environment is provisioning and it's getting ready to run. While that happens, let me show you what's under the hood. >> Logan opens a window and highlights several lines of prompts. >> In the agents.md file, you can see we created a specific set of skills for the agent to produce the show. research to pull top tech stories from hacker news, script writing to compose the show script, TTS generation gen to generate multiple speaker voices, music generation with LIA to create dynamic background music and audio mixing to bring it all together. Finally, it generates metadata, uses nano banana for cover art, and saves a complete readyto stream MP3 in its sandbox environment. I didn't write any orchestration logic. I simply define the skills and tools in Markdown files and the agent does the rest. Honestly, it feels like the hottest new programming language is Markdown, and I'm here for it. Since the agent is taking care of everything, it's going to take a few minutes. Let's switch to a tab where the agent finished and left us a nice summary of its work. Logan opens the new tab, then scrolls through the various summaries. >> Research, script, speech, music, mixing, and image generation completed, all from one API call. The playground in ASU is great to see what Gemini models and agents can do. But the real magic begins when you bring this to real apps. And to show you that magic, here's Paige. Paige Bailey comes out as an AI generated speaker name card shows Paige sipping a beverage beside a boom box. >> Awesome. Thank you, Logan. Google AI Studio is the fastest path from prompt to app. So, let's dive in. I've always wanted to build my own personalized radio station to get the latest news that I care about. And it's not just because Logan just showed us what his AI talk radio agent can do. Okay, maybe Earlier today, I took the AI radio show agent and wrapped it in a real app with just a few prompts. Since I'm also a fan of Hacker News, let's go ahead and give it a try. Just like Logan showed in the playground, my app is calling a managed agent on the Gemini API. The agent spins up in a sandboxed environment, reads through the skills, makes a plan, and then assembles the complete episode. This is going to take a minute to generate, so I'll switch over to where I ran it a little bit earlier today. And let's hear what HackerNews has to say. >> If you can go from writing 200 lines of code a day to 2,000 with an AI agent, are you actually a better engineer or are you just vibing? Good afternoon. >> Awesome. Pretty cool, right? Excellent. As you can see, the agent generated the MP3 for the show and created metadata to display the script. So now I can even jump to different speakers in the show. >> A new one. And for the accessibility community, it's massive. You can >> This is so cool. The app defines the UX and then the agent creates the content and metadata needed for it on the fly. And I love this fun cover art that was generated with Nano Banana. So Logan showed you how the playground lets you experiment. And what I just showed you is how build lets you ship. And speaking of shipping, let's go ahead and deploy my app so I can share it with the world. Deploying from AI Studio to Cloud Run is just a few clicks. You can either pick an existing Google Cloud project or if you don't have one, AI Studio will generate one for you. And starting today, brand new users can deploy to a live URL on Cloud Run instantly. No credit card required. Awesome. Now, while we're waiting, Cloudr Run is just our latest integration. We recently rebuilt our coding agent with anti-gravity and added Firebase and Fire Store support so you can easily build realworld full stack apps backed with databases and OOTH. And today we're officially announcing support for Google Workspace. Yeah. Clap clap clap. You can just prompt to connect apps like docs and Gmail and calendar and stay completely in the flow. And if you combine that with multi-hat web search and image generation with nano banana, we are completely redefining what you can build with agents. And oh awesome. It looks like my app has deployed. So you can see uh right here the status, the live URL and even unpublish right here without leaving AI studio for a second. That really was pretty magical. Awesome. I love that I can deploy my web app in minutes, but what good is a radio show if I can't listen on the go? So, starting today, you can go from an idea to an Android app directly in AI Studio. No software to install, no SDKs to manage, and no local environment needed to start. Yes. Hallelujah. Awesome. Yeah. To start, all you have to do is select build an Android app and start prompting. So, let's switch over to one I baked a little bit earlier today. I used the same exact prompts, but this time ask the app to build an Android app. And notice the app is built entirely in Cotlin. you. Woo. Yeah, you can even preview your app with the Android emulator directly in AI Studio. And when I'm ready to start testing on my own device, I simply plug in my phone to install it. We're also adding support for Google Play Store publishing in AI Studio. I just Yeah, keep clapping. I just open publish, connect my Play developer account, and push the app straight to the test track in minutes. And after that, it's ready to install on my phone without ever leaving AI Studio. You can start testing it on your own device today, and sharing with trusted testers will roll out later this summer. >> Paige walks across the stage. >> But, you know, speaking of apps, great ideas don't just happen at your desk. And that's why we're bringing the AI Studio Build experience to your pocket with our brand new mobile app. Woo! Yeah, the app rolls out in a few weeks and you can pre-register today. With managed agents, you get serious power without the complex setup. But if you want total programmatic control to customize and deploy on your own terms, today we're launching the anti-gravity SDK. Yeah, it gives developers and researchers the exact same agent harness optimized for Gemini, but with the ultimate flexibility to run it wherever and however you want. And now, of course, building a great agent isn't just a single event. It's an entire life cycle from that first spark of an idea to a deployed application. And while Google AI Studio is the fastest path from prompt to app, as your team evolves, moving to a local development platform can help with faster iterations. And moving Y's projects used to mean copying files, losing context, and completely rebuilding state. But not anymore. Today, we're launching a one-click export to anti-gravity. And we aren't just sending snippets. We are porting the complete file system and every single ounce of context so you can pick up exactly where you left off in AI Studio. And now to show you what we're cooking in anti-gravity itself. Please welcome Ancho. Paige leaves the stage. Anchel Roman strolls out. An AI generated speaker name card shows Anchel sitting at top a mountain working on a jigsaw puzzle on a small table. All right, let's talk more about agents, shall we? We aren't just making it easier to build agents. We're making it easier to build with agents. First you can kick off multiple agents in multiple projects simultaneously. One agent vodes your marketing website. Another gener architecture operating in multiple work trees collaborating without collisions. For coding tasks, you can use anti-gravity 2.0 alongside your IDE of choice. But we all know there is much more to do than just code. Anti-gravity 2.0 now is your mission control for orchestrating agents for all sorts of tasks. Now, we've all seen single agents get overwhelmed by massive tasks. That's why today we're introducing dynamic sub agents. Now, your agent can spin up specialized helpers like a QA or data science sub agent that work in parallel to get the job done faster and more effectively. Now, that might sound impressive, but what if you want the process to run autonomously? For example, you might have tasks that you want agents to repeat on a predefined schedule. Today, we are making that a reality with scheduled tasks. You can now use standard chron agents on autopilot. So, tell your agent to summarize pending PRs every morning or monitor your cloud health every hour. And this is just the beginning of making your agents truly proactive. To show you anti-gravity 2.0 in action, let's invite Kevin to stage. >> Kevin who steps up to the desk. An AI generated speaker name card shows Kevin painting a large 3D cursor. >> Thanks. Let's give anti-gravity 2.0 a challenge. As you heard from Josh at the beginning, the momentum we're seeing behind Gemma 4 is absolutely incredible. Democratizing AI means allowing anyone to take open models like Gemma 4 and fine-tune them. That meant complex pipeline wrangling by ML engineers. Not anymore. With anti-gravity, anyone can do it. Let me show you. The other day, my CI pipeline broke, which gave me this idea. Why not have an LLM automatically selfheal the pipeline? I just feed it the stack trace with the build error and say, "Fix this." to generate a remediation bash command. Anel, can you explain why I get the setup? >> Yeah. So, the challenge here is that LMS as we know have trained to be conversational. So, they bury the bash command in paragraphs of explanation. In fact, when we asked for a simple git workree command earlier, the response came back with two separate code blocks, a conversational introduction, and a detailed breakdown explaining how to choose branch names. But all Kevin cares about is the command. And so what he needs to do is fine-tune Gemma 4 to not return the fluff and just give him the command. >> Exactly. So rather than watch me type, let's go ahead and use the mic powered by our latest audio understanding capabilities. >> Kevin activates the computer's mic and allows it to access. >> Okay. So I want to fine-tune Gemma 4 to directly give me a bash command response with no additional fluff so that I can use the response directly in my CI pipeline. I have a data set of prompt to bash commands. Write the training and eval code for doing a Laura fine-tuning over this data set as well as the code to deploy my custom GM4 model on my laptop. So as Kevin runs that, I want to highlight a couple of things. He was able to just use his voice and the audio model could recognize terms like Laura. It seems like the agent is doing some research and has generated an entire implementation plan for the training code. Okay, so I approved the implementation plan and now it's writing a bit of code to make this happen. Once this is done, I'm going to take this and push it to my repo. And because we're training a model, I'll need a GPU enabled cloud machine. >> Great. So, while Kevin does that, you just saw anti-gravity 2.0, our new standalone app that is your mission control for orchestrating agents. But we know that for a lot of developers, the real magic happens in the terminal. So that's why we're thrilled to announce the anti-gravity CLI. It's a lightweight way to spin up these same anti-gravity agents right from the terminal. So it gives you the exact same harness and same models, but with a product experience that is tailored for the command line. It adapts entirely to you, your themes, your workflows, your key bindings. Kevin, we're good to go. We are all set. So let's pick up where I left off. As you can see, I've sshed into my GPU enabled VM and I'm going to pull down the code from main that we just wrote in anti-gravity 2.0. I've already have the anti-gravity CLI installed on this machine and it's running in a separate tab. No guey necessary. >> Yeah. So we know that guey versus CLI is mostly a matter of preference. But since Kevin's already in the terminal SSH to a machine, the CLI is perfect. So, I just asked the agent to start the training job that we just wrote. Um, and to give you a little bit of context, this is a lightweight Laura fine-tune, but even so, it's going to take some time before we get results. So, while this is cooking, let's just take a step back for a minute. In minutes, we used anti-gravity to go from no code to fine-tuning Gemma 4 on our own custom use case. And if all that seems a little bit too simple, that is the point. Anti-gravity makes building something of this scale not just possible, but way easier than you'd expect. All right, now that it's been a few moments, um I'm going to use the CLI to do a quick sanity check on the first few steps to make sure the loss is trending in the right direction and everything is healthy. I'll head over to my agent and just ask how is the training run going? Is everything healthy? And as you can see, it's going to look through some of the output files, actually read the code that it wrote in 2.0, know tell us that it needs to look at a couple of the log files and hopefully should give us some indication of how training is going. Um looking like it's looking through the task logs and you see that sleep 10 it looks like training has not progressed long enough for us to actually get meaningful results. So let's come back in a couple of seconds. >> Sure. Okay. Okay. So, if we have some time, let's uh ask Gemini for a joke uh using the the brand newbtw we'll get the joke anyway. >> Let's get the joke. >> Why not? We have some time. >> So, I use SLBTW, which in the same conversation will effectively fork. And now you can see this amazing joke. Why did the neural network refuse to run on the TPU? Because it heard the TPU was always a bit too tensor. I I think we'll give it we'll give it a five out of 10. We'll talk to the Gemini team about humor. Um but let's let's scroll back up because I think we saw some of the the training results. >> Um okay, so it looks like it did about 1%. Um says the gradient norm is stable. So looks like things are going well. I imagine this will probably take some time to finish. >> Yeah, I mean so the loss is turning down nicely. The run is perfectly healthy, but we'll fast forward a bit um and keep this one cooking in the background. Sure. Um, in fact, I actually kicked off the same exact run earlier. I let it run for a few hours. So, I'm just going to use that checkpoint and ask my agent to deploy that resulting model to my laptop so that we can all see that this does indeed work. So, I'll head back over to anti-gravity 2.0. I'm going to start a new conversation and I'm just going to instruct it to run the playground using a fine the fine-tuned model that is specified at this particular path from earlier. Uh, as you can see, it is running the server. And if we click over here, you can see the logs. Looks like it's running this particular model. Um, it's also starting the client. And if I just go ahead and click on the link that it's telling me to click, uh, and try that same exact prompt in the same playground that you saw earlier, we now see a fine-tune where we just get the command, none of the additional fluff. So, we just fine-tuned Gemma 4 on stage, which is pretty remarkable. And this really is the new reality of building your workflows across surfaces you already use. Awesome. Thanks, Kevin. Thank you. >> Kevin leaves the stage. Anchill smiles at the audience. >> Wrapping up on the anti-gravity CLI. We are unifying on anti-gravity as the only platform you need for agent first development. So, we took everything we learned from how you use Gemini CLI and rolled those insights into the anti-gravity CLI as well. So, starting today, the anti-gravity CLI is available to all Gemini CLI users. We've published the migration guides to help you port your custom skills over. And as always, send us your feedback as you get started. You'll now get the same harness as anti-gravity 2.0 and a unified agentic experience across all of your surfaces. All right, let's move along. Anti-gravity is completely stack agnostic, letting you use the tools you love with zero vendor lockin. However, if you are building in the Google ecosystem, we've made it easy with oneclick setup across Android, Firebase, and building on the web. >> A checklist reads, select plugins to enable Android, Firebase, Chrome. We're also making it easier to push agents into highly specialized fields by introducing domain specific skills bundles. So, starting today, our first release is a new science skill bundle, equipping your agents with the specific primitives needed to accelerate health, biology, and scientific research workloads. So everything you just saw is incredible for individual developers, but we are not stopping there. So today for enterprises, we're allowing anti-gravity to be connected directly to your Google Cloud projects, applying the same enterprise terms that you expect. And for our existing Gemini enterprise customers, you'll soon see anti-gravity rolled out in the coming months. From cutting edge research to standard enterprise apps, whether you're one developer with an idea Here to show us how that's done, please welcome Florina and Otter. >> Angel leaves the stage. Text appears. Up next, Florina Montinescu and Ut Fernando take the stage together as an AI generated speaker name card shows them biking together beside a body of water. They head to the computer desk and stand side by side. No matter where you are in your development journey, we want to make it fast and easy to build highquality Cotlin Android apps. As Paige showed you earlier, native Android development is fully supported in AI Studio. Today, we're also bringing official Android support to anti-gravity, so you can more easily deliver the most performant experiences for your users, no matter where they are or what kind of devices they have. We want to ensure you have access to the latest innovations on the Android platform and that building for the newer form factors is frictionless. >> Okay, we'll have a few demos that we need to cover in the next few minutes, but not that much time. So, let's rely on our latest skill, time travel. Travel. Travel. We'll use it every once in a while to jump into the future and check out the results. Now, for the demos, we've created a travel app. You can use it to view all of your travel plans in one place or transcribe audio diary entries using our latest ondevice model, Gemini Nano 4. >> So, using anti-gravity, let's show you how you can build an augmented experience today for display glasses. Now, setting up the environment without Android Studio is kind of a pain, but now anti-gravity has our new Android CLI built in. Now that it's stable, it comes with a set of tools for agents to handle everything from downloading the SDK to creating a project and running your app on devices, all much more easily and efficiently. Okay, Flo, for this app, we've heard from our users that they really wish there was an easier way to view important travel info when they're on the go. So, let's try this out. In a single prompt, let's add the capability to display updated flight and travel information directly on the glasses, which could be pretty useful for users at the airport with their hands full. >> Building for display glasses is new, so most LLMs don't know how to do this yet. So, as part of Android CLI, we're giving our models access to the latest information with two key resources. First, the Android knowledge base. It's our specialized data source that enables agents to search and fetch the latest developer guidance. And second, we open sourced Android skills to help LLMs understand and execute best practices. For example, you've told us that building support for edgetoedge, migrating from XML to compose or to jetack navigation 3 is time consuming. So, we made sure to add skills to help with just that. We saw in our internal testing, agents leveraging Android CLI, skills, and knowledge base use about 70% fewer tokens and complete tasks up to three times faster. Now, Android CLI also provides access to the powerful capabilities of Android Studio, like finding usages and declarations, analyzing files for issues, and looking up the latest info on dependencies. By running Android Studio alongside anti-gravity, the agent can leverage these capabilities to run tasks much more quickly and efficiently, all while under your control. Okay, let's jump forward in time to see what the agent did to build our new glasses UI. Okay, so the agent's done its thing and I can see that it used the recommended skill to build the UI using Jetack Compose Glimmer, which is part of the Android XR SDK. If I scroll a little further, I can see it's also using Android Studio to look up versions of dependencies it needs. And then a little further, it's also analyzing files that it's modified for issues. Finally, it deployed the app to the emulator so that I can try it out myself and see flight information, hotel information, and more. We're just scratching the surface of the Android capabilities in anti-gravity. How about adding a helpful AI summary of the entire trip and asking anti-gravity to give us before and after screenshots of the app? Okay. So, switching to an earlier result, I can see that the agent has implemented hybrid mode by using H Firebase AI logic and then which is great because that way the app only uses the cloud device model as a fallback when an ondevice model isn't available. >> Okay, but did it actually work? >> That's a good question. So if I scroll a little further down, I can see the before and after screenshots that we asked for with the feature implemented. I can also see that the agent used the Android CLI to deploy the app, navigate the UI, and take those screenshots. So yeah, I think it worked. >> So these screenshots are nice and all, but we need to see how this looks across multiple configurations and whether that AI summary really works offline. Anti-gravity enables you to orchestrate your agents, but the ideal setup is to use anti-gravity alongside Android Studio. >> That's right. And you can easily transition to Android Studio at any time and get that production grade polish. For example, with the agent, I can ask open the compos previews for the home screen in Android Studio and then pixels percolating. There it is. A window labeled home screen multi- preview pops open with current screenshots. >> We can see how the UI looks across multiple multiple themes and also how it adapts to different screen sizes. And then if there's anything we need to tweak, we just go to AI actions. Now we need a real phone to test the device AI capabilities. >> I agree. In Android Studio, I can deploy to a number of real Android devices, whether I own them or not. For example, here's the app running on a real Samsung Galaxy S26 Ultra, which is coming to Android device streaming a little later this summer. While the app device is in airplane mode, I can easily check that the AI trip summary feature uses the ondevice Gemini Nano. A heading labeled trip insights expands on the smartphone screen. >> All right. Now, a high quality app needs to be performant. Since I'm in Android Studio, I can view the app quality insights window. And I can see that my app has been crashing too much on produ in production. And on top of that, we also want to be ready for some of the memory limit changes for performance optimization in Android 17. So, let's get that fixed. going to go to the agent and say, "Okay, Gemini, fix and analyze my app's optimization and performance." The agent knows that one of the main ways an app can improve its performance is with R8. It streamlines your app by removing unused code and resources and rewrites bite code to optimize runtime performance. Effective RA configurations lead to fewer ANRs, reduced app size, and quicker startup times. Let's use that time machine and jump to the results. Okay, the agent's done its thing and I can see that it's used the R8 analyzer skill to actually recommend changes to the build configuration to enable R8's full mode. That's great. So if I scroll down a little further, I can see that it's using the new R8 configuration analyzer, which means in the final report, great, I have my updated optimization, offiscation, and shrinking scores. Now to achieve these scores, the agent has audited and suggested updates to my keep rules to maximize the effectiveness of R8. Previously, almost none of the app's code was optimized. With these changes, R8 was able to optimize almost all of the code. It's now faster and smaller, all from a single prompt. >> This app is getting better and better. But let's add one more feature before we ship it. We need deep links to maximize engagement and ensure that our users are taken directly to their upcoming trip right inside the app. So, if I click on this link in the booking confirmation, this doesn't currently go anywhere. In the latest version of Android Studio, we have a dedicated applinks assistant to help you implement deep links in your app. So if I go to tools, applinks assistant, open the URL mapping editor, and then I have my host jetpack or demo, then I can select which activity to handle um uh my intent, my deep links, and then I pass a sample sample URL to process. And that's it. AI is doing its job now. >> All right, you all know the drill. Let's jump over to the results. Okay, so here I can see that the agent has analyzed that URL along with the app's code to recommend a custom implementation plan. And with that plan, it's implemented the exact routing logic that uses the trip data from that URL it parsed. So if I go back to that device and click on that action in the email, it goes to the personalized trip right in the app. 90s. Ship it. Ship it. Ship it. >> Are all PMs like that or just me? >> Just you. >> Okay. Well, the good news is that now you can publish updates of your app from Android Studio directly to Google Play. Check this out. All I need to do to get started is go to build and generate assigned app bundle. Then after a couple of clicks and making sure to check this new option that I want to upload to play, >> clicking create brings up a building and signing loading bar. >> I'm ready to upload a new version of my app to play. And after I click next, it's off to the internal test track so testers can test while I get the Play Store page ready. >> Okay, now one last thing. Let's give you an early sneak peek, >> like really early, >> super early, into a suite of tools we're building to easily migrate and extend your app to Android, regardless of whether your source is React Native, a web framework, or even iOS. >> So, let's say your company is looking to expand an existing iOS app to Android's over 3 billion users. to make this a matter of hours rather than weeks. We're experimenting with a new migration assistant in Android Studio. Here's how we're thinking for doing this. I go to file, new project, and then select migrate to new project. I can choose which app I want to migrate. >> Fina selects a folder in the project repository window. >> Okay. Then I can decide whether I want to let the AI make decisions for me or whether I prefer to be more involved by selecting guided migration. I can attach reference images or even some custom skills. But then instead of doing the validation myself, I can let the agent handle that for me with journeys. It creates a set of natural language instructions following our app's user journeys. And then the agent executes, evaluates, and iterates on the app. >> Just for this demo, we've selected this open-source project called Metropolis. It's a gamified companion for the public transit network in Paris. It's really fun. You get points when you track your routes and badges when you ride the full length of a transit line. >> We ran the migration earlier and then spent a bit of time on a few polished tasks like adding maps SDK support and fancier animations. Do you do you want to see? >> I want to see. Do you all want to see >> smartphone screens appear? >> Okay. So, uh here's the iOS simulator uh and the app running in the iOS simulator and then the Android app in an emulator. So, let's say that we want to take the metro number one. >> Under maps, Florina selects one on both phones. Then we press start travel and then we go to from La Def to Leon and then confirm journey. Well, look at all the progress the agent made for us. >> Both simulators spray animated confetti, read journey recorded, and gain 165 XP leveling up. >> Okay, let's see how we got here. So the Android Studios migration assistant first creates the feature mapping and only afterwards the project plan. Since the agent knows the typical iOS and Android projects, the agent looks at the Xcode storyboard to create the corresponding Android screens. It knows how to handle things like migrating strings from iOS to Android and even how to convert assets like SVGs and PDFs to vector drawables. And then when implementing the Android code, libraries like Jetack compose, broom and view models are used together with best practices like predictive back navigation. to perfect your app and make it production ready. You can use all the great agents and tools that are already within Android Studio since this is a native Android app. The migration assistant is coming to Android Studio later this year. We are also planning on providing support for migration to Cotlin multiplatform to make it even easier to maintain shared business logic for Android and iOS. By combining anti-gravity and Android CLI with the pro capabilities and production grade polish of Android Studio, we're unifying our most powerful tools for incredible productivity gains. No matter what form factor you're building for, you can bring your ideas to life easier than ever while maintaining Android best practices. Teams of all sizes and skill levels can build cuttingedge Cotlin apps and deploy them to Google Play in record time. And in this new era of development with Google, we're extending that same level of agentic coding across every platform. Now, let's hear from Yuna to tell us how this digital workforce is delivering incredible new web experiences. >> Yuna Kravitz strides out as an AI generated speaker name card shows you watering a head shaped like the Chrome Dino. >> Hello. >> Yo, waves and smiles. Text appears. # Google IO. AI agents are transforming the landscape of development everywhere. And nowhere is that transformation happening faster than on the web. Take the browser. With Gemini now built in, Chrome has evolved to put AI assistance right at your fingertips, unlocking a new era of interactivity and possibility for users and developers. It's an exciting time to experiment without barriers, to build with confidence, to finally bring to life all those ideas you've envisioned now with a clarity and a velocity that just a few months ago seemed impossible. What it means to build for the web is rapidly evolving. Today you'll see how you can harness Agentic coding tools to orchestrate all new web experiences with intelligent guidance, automated debugging, and powerful new APIs. This is the future of the web, imagined by you, and supercharged by AI. Let's dive in. When I talk about modern web capabilities, the question that I hear the most from fellow developers is, "How do I even keep up with all of these new features?" Yes, it's exciting to see the web platform accelerate with dozens of new APIs added every few months, but first you have to figure out if they'll even work for your users. And then you have to learn how to use them. Baseline solves that first problem. With 100% of web platform features now mapped, baseline is a definitive industry standard that gives you visibility into web feature availability across the major browsers. Now, for that second problem, there really hasn't been a solution to keeping up with new features until now. Today, we're launching a new tool to supercharge your agents. Modern web guidance, a comprehensive and expert vetted set of skills providing AI agents with a blueprint for modern web features. Building on our experience with baseline, it ensures that your agent isn't using yesterday's technology, but implementing the latest web platform features and most recent Chrome innovations compatible with your baseline target. Modern web guidance empowers you to stay in the flow of experimenting and iterating while the agent handles the implementation. You can target a specific baseline version and the agent will constrain its suggestions. The guidance even contains fallback solutions and alternatives for the newest platform features that don't have wide browser support yet. to show you how powerful and easy it is to use modern web guidance with your coding agent. Let's welcome Matias to the demo desk. >> Matias Ror walks to the desk as an AI generated speaker name card shows Matias photographing a sunrise. >> Thanks Yuna. Here's something I've been working on. Dino Run, our Chrome Dino inspired car site built in React using CSS scroll driven animations. Our site unfolds as a fluid narrative experience and after that leads us into the model configurator. Let's check it out. >> A green car with adjustable features appears in his browser. >> Give it a second to load. All right. So, what color should we pick? You what do you think about this green? Or should we rather go for ceruan? >> Now, Matias, this looks amazing, but you know what? all of these menus, options, and sliders. If I was really using this, I'd want a shortcut, like a browser agent to do the heavy lifting and just help me build a car to fit my needs. So, you know what I'm going to ask? Can we make this work seamlessly with agents? >> You know, a few months ago, that would be a really hard question to answer, but not anymore. With WebMCP, we can now make this web page agent ready in minutes. This proposed browser standard lets you expose web capabilities to browser based agents. In other words, it lets you tell agents how and where to interact with your site, making those interactions more precise and reliable. >> Now, we could implement this ourselves on this stage, but with modern web guidance, our agent has the skills to do it on its own. All we need to do is ask. Matias, could you give it a try? >> Absolutely. Let me open anti-gravity where I can just prompt please implement webmc tools for the car configurator on this page and off we go. Modern web guidance is a set of textbased skills that are internally tested, benchmark proven and token efficient. So for web development tasks, the jump in pass rates in guided compared to unguided coding is an average of 37 percentage points. And you can even read the markdown files yourself. Oh, let's go ahead and allow this all. There we go. Let me pull up an example for a guide. This skill contains knowledge about WebMCP so your coding agent can jumpst start the development process and handle the groundwork with confidence giving you a solid foundation to test and refine. Here you can see how it explains that web MCP is made up of JavaScript functions that are exposed to agents through an API. We've made accessing and implementing modern web guidance as easy as possible in anti-gravity. It's easy to install with one click during onboarding or later on from settings. But if anti-gravity is not your coding agent of choice, you can install it as a ready-made package of skills for other coding tools. And because the guidance is explaining core platform features, it adapts to your framework of choice, no matter if it's Angular, React, or any other. All right, let's check on progress in anti-gravity. All right, looks like it's finished. So I am switching back to Chrome and I'm going to click Ask Gemini in the top right. This brings up a prototype of Gemini in Chrome with experimental webmc support. Since this is still in active development, the final version may differ from what we will see today. And while this demo is Chrome specific for now, once WebMCP stabilizes, these tools will be compatible with any browser based agent that supports WebMP. Yuna, let's prompt away. >> Okay, let's have some fun. Can you prompt in Gemini and Chrome? Configure the ultimate party car. I want immersive audio and some interior lighting would be great, plus enhanced visibility for night driving to keep me safe. And I don't want to go too crazy, so keep it under $40,000, but give me as many add-ons as you can under that. >> All right, this sounds like an excellent build. I'm briefly double-checking the prompt, and here we go. Now, Gemini in Chrome creates an auto browse plan and briefly passes it back to me for confirmation. I'll briefly approve the task, and then off it goes. With WebMCP, Gemini and Chrome and other browser agents have one more tool in their box to interact with a website. In test runs with this site, anti-gravity has implemented an imperative tool called update car configuration with all configuration options listed in the schema definition. Now, Gemini and Chrome can use this specific targeted web MCP tool for the job. And through modern web guidance, we were able to implement it and make this app ready for the agentic web in no time. Yuna, your build has finished, meanwhile. What do you think? >> Let's talk about the color later on, but this is so much easier than finding all of the right options myself. Modern web guidance with support for over 100 use cases for dozens of the latest features is available for early preview today. The purple vehicle gets replaced with text reading modern web guidance available today in early preview. I'm also excited to announce that the experimental web MCP API will enter origin trial starting in Chrome 149 and Gemini and Chrome will soon support your web MCP tools building on the active experiments we're running with our ecosystem partners. The logos for 10 businesses appear on the screen. It's >> exciting. Okay, so we've added some new features to our website, but could we have the coding agent now test this like a real user would? >> Yes, we could. I think by now most folks here have tried building with agents. They've gotten a lot better, but you still don't always get what you prompted for, and that's because your coding agent can't actually see what it's coding. That all changes with the new Chrome Dev Tools for agents. Thanks to in Go ahead. Thanks to incredible feedback from our early preview, we've evolved it to do even more. Now your agent can finally see how the code it has written performs at runtime. With the FCP server, the CLI, and a set of tailored skills, it finally creates a closed feedback loop for agents. That's phenomenal for building, validating, and debugging. To see its capabilities, let's go back to anti-gravity where again prompt. Please check the web MCP implementation with Lighthouse. Now, what will happen in a second is you'll see Chrome popping up loading our page. Basically the same thing that would happen if you render Lighthouse audit through the DevTools UI. But I think I need to allow one more tool call. There we go. There is Lighthouse running. And Lighthouse runs with a new agentic browsing category which enables Lighthouse to run a comprehensive health check for the Agentic web. It validates your web MCP tool registrations and ensures your forums have the declarative metadata that agents need to be successful. >> It also verifies your LLM.txt txt file, a new standard for giving models a clear map of your site's content, and it's re-evaluating a lot of the familiar Lighthouse accessibility audits. Because most agents navigate the web using the accessibility tree, every Arya role or label that you optimize doesn't just help a human, it makes your site more actionable for an agent, too. When an issue is surfaced through Lighthouse, you don't need to be the human clipboard anymore, copying an error from DevTools, passing it into a chat, and hoping that the agent guesses the right fix. The agent reads the report for itself, attempts a solution, and can rerun audits to see if that solution worked. It's a supercharged workflow that automates the busy work, leaving you free to spend your time innovating. If you want to give it a try, Chrome DevTools for agents is available today for anti-gravity and more than 20 other coding agents of your choice. >> Okay, Matias, I think that we have one more surprise hidden away in our car configurator. >> Yeah, we do. Let's go back to your party car running in canvas. Let me switch to the interior view. Do you see the screen in the center? It's actually interactive. I can click through to the settings screen of our car and adjust the ambient lighting just by pulling the sliders. And you know what? If we do a quick inspect in DevTools and select those elements, the full display UI is actually native. Those are native HTML elements rendered into the canvas. Woo! I know what some of you out there might be thinking. HTML elements inside of a canvas. Well, that shouldn't be possible. >> The new HTML and canvas API now does the impossible. With this, you can now integrate real DOM elements directly into your canvas environment. Now, you no longer have to choose between visually complex and stunning or interactable and accessible. >> Exactly. Because it's part of the DOM, every element in your canvas is now searchable, accessible, selectable, translatable, and interacts with your built-in browser features like autofill. And that creates some really interesting possibilities. >> Yeah. For example, you can style them just like any other DOM element by adding a little class like this one. >> Cutouts of their heads are used for the ambient light adjustment slider design. We've already seen the community build amazing demos with WebGL textures, 3D interfaces, and entirely new modalities for interaction with real DOM content. HTML and Canvas is an origin trial now, so you can test, experiment, and build. Now, I know a lot of web developers are really excited to try this out, and that's because this is more than just a new feature. Alongside everything else you've seen today, it represents a fundamental shift in how we build, who can build, and even what we can build. In this new era of agentic web development, there's never been a more inspiring time to create with tools like modern web guidance to help you quickly and reliably implement new features like webmc HTML and canvas to truly expand accessible, interactive, creative new UI on the web and Chrome dev tools for agents to autonomously debug and test performance. The future of the web is imagined by you and supercharged by AI. But this isn't just about the web. It's about how we build everything. So to bring it all together, please welcome back Josh. Josh Woodward returns to the stage. Yuna and Matias leave together. >> All right. What a show. From Google AI Studio to anti-gravity, from Android to the open web. Today, you've seen agents completely transform the developer experience. The common thread across all of this is simple. You focus on the big idea and the agents can do the heavy lifting. And it's never been a better time to build. We hope you'll take what you've seen today and push it to the limits. To get you started, we've created the ultimate platform to make an impact. We are officially launching the Build with Gemini X-P Prize Hackathon. This global hackathon is going to offer up $2 million in prizes for builders who create apps that solve actual real world challenges. And the premise is simple. Pick a problem worth solving. Build with Gemini and let's all try to positively impact the lives of a billion people. To build at that scale, you're going to need some serious power. And as you heard this morning, we announced a new $100 per month Google AI Ultra plan alongside a simplified pricing structure based on your feedback. And to keep those agents humming over the long weekend, we're giving all Ultra subscribers $100 in bonus credits today. You can claim the offer right in the anti-gravity app, and the credits will kick in when you hit your limits. So, let's get going. You can visit io.google Google for the live stream sessions and all the on demand content we'll be releasing over the next few days. Thanks for coming and have an amazing IO. >> Josh leaves the stage and the view floats up over the crowd. Text # Google IO results for illustrative purposes. Sequences shortened and screen images simulated throughout. Some products featured may require a paid subscription, may only be available to users 18 plus, may require a setup and/or compatibility and availability may vary. Some products featured may be prototypes are still in development. Any final product details may differ. Check responses for accuracy.