Build Next-Gen AI Experiences with Google AI Studio and Google Antigravity

[MUSIC PLAYING] JOANA CARRASQUEIRA: Welcome, everyone. Hi. How are you doing? [APPLAUSE] OK. OK. That's more like it. How's the energy in the room? I am Joana Carrasqueira. I'm a Group Product Manager with the 1P team here at Google. And you might have seen Anshul earlier this morning. ANSHUL RAMACHANDRAN: I'm Anshul. I work in Antigravity. JOANA CARRASQUEIRA: Awesome. And we're going to tell you all about what's new in Google AI Studio and Antigravity. Just right before us, Ammaar and Paige, they walked you through all the experience in Google AI Studio. And now we're going to go one step deeper. Does that work for you? ANSHUL RAMACHANDRAN: Good. Sounds like a plan. I think you're going to get us started? JOANA CARRASQUEIRA: Let's get started. Are you guys ready? OK, awesome. So at Google DeepMind, our goal is to build AI responsibly, that benefits humanity. And everything that we do is always with this goal in mind. We always want to make sure that everything that we build, from the models, from the products, from the services, they always benefit everyone. And everyone has a positive experience with our technology, regardless of where you are at, your language, anything. And with this very ambitious goal, we do so many different things at DeepMind. And we bring all the best and greatest to the market. And we now live in a very special time where you can build anything. And in these times in which we can build anything, the rules of software development have been written-- rewritten. The what and the how are changing dramatically. And how you react truly matters. And this is why at Google, we built a stack. And we built an ecosystem that allows you to have tools for every different type of experiences that you're trying to build. And we have this very robust ecosystem that allows you to go from exploration to development to Cloud deployment very seamlessly. And our products are always powered by the latest and greatest model innovation. So Google AI Studio is becoming a one-stop shop where you can find the models in the playground, agents, the build experience. And I'm going to show you that in just a second. But now we're building more of these product integrations. You've heard about the Google Workspace integration coming to AI Studio, the Cloud run into Antigravity. So more and more, the Google ecosystem is being built in a way that provides you with a solution regardless of what you're trying to build. And we typically say that Google AI Studio truly is the fastest way to go from prompt to app. And it truly is like that. And I'm going to show that to you in a second. This is what Google AI Studio really is and what it looks right now. And it has evolved so much. Just a few years ago, we were showing Google AI Studio. And it has evolved from being a playground where you could try and test the models and the different capabilities, to now being a place where you can try not just your ideas. But you can also deploy your apps into Cloud Run. We've been bringing more of these product integrations, like I mentioned before, with Google Workspace, but also with Cloud Run for a one-click deployment. And more and more, we are building a product that is very robust and is tailored and catered to you so you're able to build anything that you want. And the team is building at relentless, relentless pace. When we say this, honestly, you have to trust me. Everything happens so, so fast. We're shipping new things almost every day, it feels like. And I just wanted to bring this list of all the things that we've announced here at I/O, just so you can see the sheer amount of work that goes behind these products that we all love and use on a daily basis, from the mobile app, to direct deploy, to the Play Store, to the Google Workspace integration, to the Google Antigravity coding agent. There's just so, so many things that are coming every single day to our products. And a lot of that is a result of your feedback, the developer community, that tells us exactly what to improve and how we should keep improving the experience for you. But better than showing it to you on a screen, why don't I go on my laptop and show it to you, how it actually works in the playground? Does that work? [CHEERS] OK. OK. So let's do this. If we go to the playground, if you go to the laptop-- OK, there we go. So this is what Google AI Studio looks like. And Paige and Ammaar, they gave you a very in-depth overview. So I'm not going to bother you with that. But I just wanted to show you the playground where you have our models, our new agents. And we have different agents coming to Google AI Studio. For example, one that I really like is the Data Analyst. There's a bunch of others that you can try, the Customer Support, the Document Processor, so many different agents for whatever use case you're trying to do. Now this morning, you might have seen. Josh showed Louie Cinnamon on stage. Do you remember? Yes? OK. So I was really inspired by his story and how he has to find a pet hotel in order to keep his pets safe while he's on holiday. So I had an idea. Why don't I actually build a new business? And so I thought, I'm going to open a pet shop. And I'm going to have a place where we can have all the pets coming in and spend time at the hotel. However, I don't know anything about the pet shop business. Do you guys know anything at all? Anyone that can help me? No? OK. Good thing that we have agents in AI Studio, because I am going to ask for their help. And so I already have a prompt which I also refined in Gemini app. And it's a prompt that is-- I'm going to ask the Research Agent to do a bit of market research to see if my idea is actually feasible here in the Bay Area. So my prompt I'm going to ask, you are a premier retail market research analyst specializing in the San Francisco Bay area and Silicon Valley commercial landscapes. I'm planning to open an independent brick and mortar pet shop in the region. I want a comprehensive deep research, market analysis. And then I specified everything that I wanted the Research Agent to find-- target demographics, micro markets, correlate pet ownership with demographic data, inventory strategy, recommendations, and everything. So I'm going to just basically paste my prompt into AI Studio. And I ran this early today because Wi-Fi and live demos. And I have here-- and this is basically what it looks like. I gave it my prompt. I saw the thoughts. I approved the research plan. It has all the specifications that I asked it to search for. And now it has given me a very, very in-depth report with step by step exactly of what I should do. Now this is just one example of the Deep Research Agent. But if you come here to this section, you have the Max preview, which is better for overnight tasks. But you also have other types of agents, so you can pick and choose. And everything is just available within the same section of your model selection. Now let's say you want a very quick analysis instead of something that is actually very, very time-consuming. You come here to System Instructions. You pick the system instructions or the persona you want the model to impersonate. And you add a description. You can also add constraints. And they will do the research for you. So I click this one. I'm just going to basically paste the same prompt. And I will just let it run. And it's going to run for a while, which is absolutely fine. So while the agent is going to show its research, I keep thinking, OK, now I've done my research. It's a good opportunity to open my pet shop in the Bay Area. What should I do next? Find a place. OK. That's a great one. But how will people know that I have a pet shop? A website. Someone said website. OK, wonderful, wonderful idea. And here is my website. Let's say that I went back home, started working on my website, on my inventory. And this is exactly what it looks like. What do you guys think? Now, fun fact, all these pets that you see on screen are actually our pets. So you can see Josh's dogs. You can see my little Jon. That's my pet. So this is all built with pets from Google. So I think it's the most wholesome moment of I/O. And this is what the website looks like, very, very basic. What do I need now in order to improve my website? Come on, guys. Feedback! We're developing a new product. We're developing a new business. We need feedback. How can I get feedback very, very quickly? OK. I'm going to show you. So this is quite a new app that is coming to the Google AI Studio app section. It's coming very, very soon. And this is what it looks like. This is probably one of my very, very favorites. You can remix this app. And what it does is an agentic focus group. So it uses different agents to impersonate different types of user segments and audiences that you may want to try to attract with your own business. So this is extremely useful. If you have an online business, if you are a startup, if you are a builder trying to identify your niche audience, you can basically, with a remix, just adjust whatever persona might be within your business. For the purpose of this demo, I'm just going to keep a Skeptical CFO, a UX Lead, a Status Seeker, so just very generic personas. But I'm going to get my website. And I'm going to run it with my agent focus groups. And now while it's analyzing, it's going to basically give me a description. What's the sentiment analysis? What do they think about my website? And so I can tailor all these different-- I can tailor what they tell me and get that feedback into the next iteration of my website. Now let's say I want to keep building my website. I already know it's really cute, but it's not the very best one. A very cool thing that I can do is come to Stitch. And Stitch is perfect. I can just say, create "a landing page for my pet shop in the Bay Area" and let it run. It will run and come up with some design examples, hopefully-- maybe not. There we go. And if I'm happy with one of the design examples that Stitch came up with, I can simply export it into AI Studio. And I can keep building with AI Studio. So when I give-- and it opens directly in the Build tab. And I can now say, build me an app with the screenshots that look like this. And it will just build a new website for me that is based off of the Stitch designs that were based on my initial idea. So that's something really cool. Now I'm thinking, OK. So I've done my market research. I have my website. I've improved my website with my agentic focus groups, and this is what it looks like. So if you go here, you can see the analysis, the sentiment, the match radar for each one of the personas. So let's see. The Busy Executive says, "While heartwarming, the current site forces a high-friction, time-consuming adoption process," so things like that. And as you can see, it just changes based on the persona that I'm choosing. OK, cool. So I did my market research, built my website, improved it with focus groups. I wasn't really happy, built something new with Stitch, exported to AI Studio, kept on building. OK. I think this business is going to be successful. What do you guys think? Oh, yeah. OK, great. Now what am I missing? Inventory. Thank you so much. Now, my inventory-- I've been doing my inventory in a basic way. I've been using a Google Sheet. And I have all the ideas, the different names of the pets, the animal type, the breed. I added pictures. But those pictures are also in a Google Drive. I have the health status. OK. It is good. But it could be a lot better. Don't you agree? Wait a minute. Google AI Studio now allows integration with Workspace. How about I also build my next inventory management dashboard with Google AI Studio? Let's do that. OK. So I go here. I go to the Build tab, and I say Create. Create a dashboard for my pet shop and allow for integration with Google Sheets for inventory management. OK. And now it's going to build. Now as you might have guessed, I have already pre-loaded it beforehand. So I'm going to show exactly what it looks like. I had given the model the exact same prompt. And I kept on iterating it just a little bit. So I improved some of the color schemes so it would be a bit more on-brand based on the feedback that I had received from my agents. And while the model is thinking, it had asked me to integrate with Google Sheets. Let me just go scroll up so I can show it to you in the history. OK. OK. Cool. So now let's give it a try, see if it works. It does. Awesome. I now have an inventory management tool completely integrated with Google Sheets-- with Google Drive. And I can see the total number of pets, the number available, adopted, their health status. I can see the asset gallery. And I can basically manage all my different animals here. If I see something that is not right, so for example, if I try to find John, John is not available. John is my cat. I want to change that. OK, move it to "Adopted." And the cool thing is that you have real-time updates, both in the app and in the Sheet. So if you change something here as well, "Needs Checkup," it will change in the app as well, which is really cool. OK. I've done everything-- research, website, improved the website, built a new one, created an inventory management. Now, my last step, what is my last step? Finding clients. OK, very cool. So I go to AI Studio again because I'm like, OK, I need a strategy. And I don't know how to build that strategy. And I'm going to ask the model, "Create a marketing tool to help me with go to market for my pet shop. Give me 10 CTAs and help me optimize ads." Cool. So now the model is going to run my next prompt and basically has just given me this result that looks like the dashboard that I want for my go to market strategy. So I can literally just come here, get the link of my website. And I am going to turn this into a growth strategy that allows me to then literally create any type of marketing assets-- my hooks, my visuals. It can integrate with Nano Banana, with Veo. You can create a brand identity, a logo. So literally, Google AI Studio is becoming a one-stop shop for startups, for builders, for small businesses. You can manage and create an entire business, as I just showed you. And now I have my growth score based on my current website. I have a friction analysis. I have SEO and metadata analysis, all my campaigns in just one place, plus my ad simulator, ready to go, ready to ship. OK. So this is what I had to show you in Google AI Studio. There are a few very cool things that I would highly recommend you to test after I/O. Agents is definitely one of them, customer support, data analysis, research, and Antigravity in AI Studio, and then some of the new focus groups that will help you take your business to the next level. And one very cool thing that we brought into AI Studio is the integration with Antigravity. And Anshul is now going to show you everything that is new with the Antigravity. ANSHUL RAMACHANDRAN: Awesome. Everyone, give Joana a hand, a round of applause. [APPLAUSE] Cool. So I'm here to tell you a little bit more about Antigravity. You've probably heard about it a little bit throughout the keynotes today morning. But I really was going to spend some time here today to give you a little bit of a thought process as to how we designed a lot of the new features in Antigravity. But the big announcement today morning was Antigravity 2.0. So when we released the Antigravity IDE back in November-- it's shockingly only been six, seven, eight months, something like that-- we introduced it primarily as an AI-powered IDE. But in the Antigravity IDE, we did have a second surface. We called it the Agent Manager. And it was kind of our view of what would the world look like in an agent-first kind of world. If you think about it, if most of the time your agent's doing most of the work and your IDE is 70, 80% not your agent, something seems off. And so we started designing what does that actually look like-- what happens when the entire UI starts focusing on the conversations you have with your agents, the management of all of your agents, and any kind of artifacts and communication that the agent does with you. And we were like, OK, we'll put it there. We won't make that the primary surface. Let's see if people actually enjoy this. This is long before these kind of agent-first kind of paradigms became popular. The short answer was yes. And so it became pretty clear to us over the last few months that what we needed to do is separate out just that agent-first surface into its own standalone application. That allows you to just deal with the agents. And if you want to code with Antigravity 2.0, you can use any IDE of your choice. But if you want to start using Antigravity for more and more tasks beyond just coding, which you all have, now you have just one place which is just truly dedicated for that. And so that's really what the end goal is, to build the agent-first platform that's built for the next era. And, OK, show of hands, how many people are a little bit tired of hearing the word "agents" today? A little bit? OK. I'm not going to make it any better. So I'm sorry. I'm going to be talking a lot about agents. But really, how do we think about working with agents in a way that you can get the most done, you can stay focused while also keeping security and permissions and all of these things top of mind? We launched a lot. So these are all brand-new things that we launched today. I think you guys found it a little bit exciting because you all took down our authentication services, for just how many people were trying to log in. It's all fixed now. So if you got some issues earlier today, try again. Apologies. But there's really a lot that we're trying to do. We're trying to improve the core agent, the primitives that it has to do more powerful work more quickly and more efficiently and better. But then we're also thinking about how the UI should actually evolve so that you're looking at agents and managing them appropriately, because really, we believe in a world where you will have n agents working together across x different projects. And that is a completely different way of working. And hopefully, that helps everyone multiply what their output looks like. And then we added a lot of nice, fun product features as well. So I'm just going to take you also to the laptop. Now I'm not going to spend a whole lot of time actually trying to build something. I'm not nearly as good with that live. And I think it's going to take a little bit more time than Joana did. There is a 20-minute demo that Kevin, my co-presenter during the dev keynote, put together that's online. That will show you every single feature in action. But what I really wanted to spend today is walking you through the Antigravity 2.0, showing a little bit of the lay of the land and some of the design decisions that we took in so that you can get the internal understanding of how we thought about building it, and therefore how you can actually get the most out of it as well. So as you can tell, I'm not sure. But anyone, does this look like an IDE? No, not really? Yeah, this is really it. This is the kind of minimal entry point to Antigravity 2.0. It's you, your agents, and the conversations. So to get started, this is a brand-new laptop. So I just started a task earlier just to be able to highlight some of the pieces here. So in the middle is your conversations with the agents, every agent, every conversation. You can have it multi-turn, all of those things. On the right, there's the side panel that has a lot of different pieces here. And I'll walk through some of them. The first part here is sub-agents. So this is something that we newly added to the product. The reality is, people are trying to do much more complex tasks with Antigravity. And if you have one agent that's trying to do all the work, you might give it a much large, vague task to go out there. And I think we've all seen LLMs get a little bit confused if it's trying to multitask itself. That's common. So what we wanted to do is make it possible for that main agent to itself determine that, hey, I have this well-scoped task that needs to get done. I should just hand that over to a sub-agent that's just focused on that one specific task. It can complete its work. And then it can report back when it's done with any kind of outputs that might be interesting for my main conversation. And why is this nice? It allows the main agent to stay focused on whatever that big, large, vague task you gave to break it down. And it allows the sub-agents to be really focused. And it also allows for parallelization. If you have n tasks that could be done at the same time, you might as well do them in parallel. Actually, let's show this in action a little bit. So I'm going to just start a new conversation. We'll open up in a new project. So I'll talk about projects in a little bit. But you can also just run your own standalone conversations if you have that idea on the side that you just want to do it. I'm going to be pretty explicit, just so you can see the sub-agents in action. But I can say, spin up a couple of sub-agents to tell me all the things launched in Google Antigravity at I/O. So hopefully, this works, and the demo gods are with us. Yeah, there we go. So what you just saw here is, really quickly, a couple of sub-agents being spun up, as I pretty explicitly asked in this made-up example. And if you actually look and if you click into them, you can kind of see what was that prompt that the main agent passed into the sub-agent to do. It will all happen in parallel. So I can back to the main agent. And as you can see, the sub-agents are reporting back to the main agent. Hey, I've just finished my task. Here's some information as you asked me to. And that main agent can now take that information and synthesize that together. So this is just a really simple example of sub-agents in action. But that's not the only place that we tried to improve the core agentic experience. The second one is, we decided-- we realized that, OK, you're starting a new project. A lot of time actually spends in commands like installing packages or doing updates. And the agents are doing all this exciting work. And then, boom, you're just waiting there for five minutes for all these packages to install or update. It turns out there's a lot of tasks like these that are just long-running commands that can actually happen asynchronously in the background. So asynchronous processes and asynchronous task management is now kind of a core primitive within Antigravity. So I want to see whether or not that actually happened in this conversation I had over here. Doo, doo, doo, let's see if it worked. Yeah, so you can actually see, running these installation commands actually happened in the background. And once it was completed, it kind of reported back that, hey, this background task, this background command got completed. So this also will allow your agents, just like sub-agents allow you parallelize, this is another core primitive that we've added to make sure that your task can get done quicker. In fact, these two things actually compound with each other. You can imagine the agent launching sub-agents in the background. It can do its work while it can continue doing its work. So actually, what was happening here, if I remember correctly, all the installation was happening while the main agent was writing the code. And so by the end, the packages were ready. The code was ready. I could actually view the app, just like that. So these are some of the core primitives that we're putting together. The third core primitive that we added to agents was this idea of hooks. So these are programmatic pieces that you can actually have your own custom script. And this is all defined in a JSON file. It's all in our documentation, where you can actually tell, hey, every time before you run this specific tool or before you stop your turn, make sure you run this script. So what's a really good example of this? Maybe after every conversation, you want the agents to run some kind of validity checks that is specific to you or your company. You can define that in a custom script. You can register that as a hook for that stop command. And the agent will make sure that that is actually always called programmatically. So this is actually a way for you to be able to inject your own custom logic into the agent's behavior. So a lot of abstract topics, but these are all ideas that we've actually tried to bring into Antigravity 2.0, so the agents are more powerful in this service. So that's some of the agent pieces. But let's talk a little bit more to the app. I was kind of going through this right-hand panel a little bit. We talked about sub-agents. We talked about background tasks. Actually, the npm run dev is still running, so you actually can see that that's up. But what we also show is what we like to call artifacts. These are essentially, if your agents are running for longer and longer by themselves, autonomously, we want to make sure that they can communicate their progress to you in ways that you understand. You can already see me go through this conversation. I was scrolling through. And I'm like, oh, what happened in this whole conversation? What happened where? It's a lot of text. And it's a lot of tool calls, and these agents are doing a lot. So that's not always the best view into knowing what the agent did. So artifacts are essentially these ideas of-- they could be markdown files. They could be code changes of, hey, this is something that I actually want you to look at that will actually tell you a little bit about what I'm doing so that you can build your trust in the agents. And you can know that the agents are going in the right directions. So for example, at the very beginning of this kind of task, it generated an implementation plan. So this was like, hey, this is what I'm planning to do, all of that. The nice thing about all these artifacts, actually, is that you can comment on them. So whether they're text artifacts, visual artifacts like screenshots or anything like that, you can leave comments. So this is one way. It's like, hey, this is my plan. This is what I want to do. If you're like, actually, no, instead of trying to battle with the agent back and forth, if there's just some small nuance you want to do, just leave inline comments. And the agent will incorporate that as it continues doing its work. So this is one way for you to guide the agent. Other things that happened, it had a walkthrough after it was all done, a task list. And of course, it also changed some files. So I can go through all the files it changed. And one thing that we did in Antigravity 2.0 that, it wasn't there in the agent manager and the IDE, is that we made it a pretty easy flow to actually review your code also inline. So write some code. If you really want to get in there in the nitty gritty, the entire review flow is now directly in Antigravity 2.0. Cool. So those are some of the kinds of things on the right-hand panel, which is really your way of investigating into that agent conversation and knowing what's going on. But as I said in the beginning, I will say "agent" a lot. Our goal is that you guys are actually running many agents in parallel across all your projects. Maybe you have multiple ideas. You have the app that Joana built. You brought it down locally into Antigravity. And you're like, actually, I have this idea and this idea and this idea. Our goal is that you should be able to fire all these agents in parallel, and it should work. So we spent a lot of time actually redesigning what does this left sidebar look like. And it might not seem like a whole lot, but there's actually a lot of ideas here. The first is this idea of projects. So historically, we had agents tied to repositories. That was the core primitive. You have a code repository and an agent that works on it. We decided to actually broaden that a little bit more. So a project is just a general purpose collection of resources that an agent can work on, and you can apply permissions on. So what do I mean by that? I have this todo-app project, so I can go to the Project Settings. You can actually connect multiple folders to the same project. So if you actually want your agents to be working across multiple repositories-- you have a microservice-based system, whatever it might be-- that's actually really easy. You don't need two agents in the two different repositories to try to work with each other. You just create a project that actually allows those particular agents to access those x repositories at the same time. And you'll be able to make edits across all of those. The other thing that we ended up doing is we ended up creating projects so that you can define settings and permissions at the project level. I think we all know that-- we probably have all seen things online of agents going wild, deleting a bunch of things that you didn't expect it to delete. And you're like, how did I end up in this world? And really, a lot of that boils down to, as your agents do more and more work, it's very natural for us to be able to just give more and more permissions to your agents. Actually, I was like, stop asking me to approve every single one of my commands, which is the default kind of behavior for all these agents. And next thing you know, all your agents can do pretty much anything on your computer. As you probably guessed from how I framed it, that's not the most secure way of going about building with agents. So what we're allowing to do with projects is for you to be able to define those permissions and settings at the project level. So that means only the agents within this project have these scopes that are permissions. So you can specify which kind of files do you actually want to allow it to access outside of the folders that you've connected it to. What are the kinds of different review policies, the terminal commands? So for example, you can say, always allow these kinds of commands. Never allow these other kinds of commands. And if you're working in repositories that you know have more critical code, you might give fewer permissions and want the agent to ask you more frequently every single time it wants to run a command. If it's just your todo-app, like here, you might just say, go wild. Build something as long as it's not rm -rf or or something like that. You'll give it a whole lot more permissions. So this is the direction that we want to go to, this core primitive of a project that we're introducing that allows you to group agents in ways that aren't just necessarily particularly tied to a single repository. It's actually tied to a whole lot more. so a lot of stuff to explore here. And all the customizations, your MCP tools, your skills, all of those things, you can attach. And you can see a breakdown of how you're using your tokens within the system prompts efficiently. Cool. All right. What else do we want to show around here? There's a bunch of other fun, little tidbits that are within the product. As you probably saw on stage during developer keynote, there's now a live audio transcription. I will say, I don't think maybe every single Google office has been well-designed for everyone talking to their computers. But that's the new reality, I guess. But I think research and talking is 5x or something faster than typing. And so sometimes, when I just have an idea, I'll just go and talk to the agent. And one of the really nice things about the audio model here is it isn't just one-to-one live transcription. It's actually using our latest audio models that will synthesize and remove all your filler words and make sure that you have cohesive thoughts so that the prompts that you're sending to the agent are actually well-formed ideas. So all my "ums" go away, which is great. Some more features that we've added-- everything that I've shown so far has you kind of working synchronously with an agent, which is still a pretty important paradigm for people working with agents. But I mean, I think you guys all saw some of the stats with Gemini 3.5 Flash earlier today. It's a workhorse of a model. We've been using internally for a couple of weeks, and it's been crazy. And on Antigravity, we've been able to do further infra optimization so that it runs about 12x faster than frontier models. So we see somewhere up to 700, 800 tokens per second being produced. So you run it, and the thoughts kind of pass before you even have time to look at it. So it's a really, really fast model, but at the same time a very, very powerful model as well in every benchmark and every single AP test that we've run internally with Googlers. And so with that increased intelligence, it really got us thinking, OK. We're going to definitely start moving to a world where people are just going to want agents to run in the background for them, asynchronously. It shouldn't require me having to kick off every single agent, every single time I have a conversation. In my day to day, I have a lot of tasks that I pretty much do on a regular cadence. Every morning, I will use an agent to make a digest of all the PRs that are open to review and what's actually high priority, what's not. all that. I like to have a digest. These are really common things that we want to have on some kind of chron schedule. You launch a new job on GCP. You want to make sure that there's no issues. And realistically, if there are any issues, I would want an agent to first do a first pass of debugging on what the issue is before I have to actually dive in. So these are all common examples of where you want an agent to be doing something on a schedule, layer that to talk about scheduled tasks-- not very clever. But that's pretty much the basic idea. The simple thing here is you just made a really simple, new scheduled task. You enter a task name, like "Daily Digest." You set the project so the agents that are spun up on the schedule will obey all the same permissions and everything that you have for that project. On a daily schedule, 9:00 AM looks good. You put in your prompt. Summarize my PRs or whatever, dot, dot, dot. And you just add the scheduled task. And it's there. It's really that simple. And so those tasks, those agents, will show up in your project as they run. You can check in whenever you want and be like, OK, that looks good. And because they're responding here, if you wanted to do follow-up questions, this is all a conversation. So just go in there. You pick up the conversation. You continue. All right. I remember having this UI, and someone's like, actually, what I really want is I don't even want to go to another panel. Can you just have a slash command? So we introduced some new slash commands. So there's slash schedule. So you can actually just subscribe in natural language. Hey, every 9:00 AM, run me a digest running this project. And we will generate that schedule task for you. So explain natural language. You don't even need a form. Forms are so, I don't know, 2025. I don't know. But we have a few others that we put here. So, often, agents will come back and they'll ask you for feedback, like the implementation plan, something like that. But sometimes you're like, hey, just keep on pushing through. Don't ask me for feedback. Just make your best decisions and get to an output when you actually think the job is terminated. That's what slash goal is for. So slash goal is like, hey, here's my goal. Just keep driving. You don't need my input. Just go for it. Slash schedule, we just talked about. Slash browser, one of the favorite features when we launched the Antigravity IDE was the fact that the agent is fully capable of spinning up a browser sub-agent of itself using the Chrome DevTools MCP to actually actuate the browser. So if you're building an app, the agent can go out, test the app itself, give you some feedback and take screen recordings. So slash browser essentially just tells your agent, hey, make sure you use your browser tools and the browser sub-agent do that. grill-me is a fun one, which is-- I don't know about you, but I sometimes write the vaguest and dumbest prompts. So grill-me just essentially is a way to, hey, first ask me clarifying questions so that we get to a good point before you actually go out and start implementing things. I probably missed, honestly, quite a few features that we introduced here in Antigravity 2.0. And I've not even talked yet about the Antigravity CLI, the Antigravity SDK, the Antigravity agent in the managed API. So we're really trying to bring this agent and this agent harness that is being co-optimized with the Gemini models to you all, in any kind of way that you find best. So if you looked at all of this and for the last 20 minutes, you were like, Anshul, I absolutely hate a GUI, bring me a terminal, don't worry. We have that covered. I just don't have time today. Try out the Antigravity CLI. And if you're a developer who's like, hey, these agents can do really powerful things. Gemini 3.5 Flash is a very powerful model. I want to start integrating that into my own projects. I want to do some kind of research on that. There's both, as I mentioned, the manage agent within the Interactions API, as well as the Antigravity SDK, which will allow you to actually take that same harness, but customize parts like the system instructions, the tools, all those customizations of the agent, and allow you to deploy that wherever you want to deploy. So a lot that's been announced in the last, I don't know, eight or so hours here at I/O. And I'm sure there's going to be even more that's going to be talked about. But I think we're roughly wrapping up on time. So I think we're going to start wrapping it up. And I'll let you guys get to the block party. JOANA CARRASQUEIRA: Yeah. Yeah. I mean, I only have two last questions for you. One is, who's ready to build? [CHEERS] Yes. Yes. And who's ready to go to the block party? [CHEERS] OK. OK. ANSHUL RAMACHANDRAN: That was louder. JOANA CARRASQUEIRA: That was louder. ANSHUL RAMACHANDRAN: Yeah, I don't know about that. We'll have to work on that one. JOANA CARRASQUEIRA: Yeah. Yeah. I think we've seen so many cool things here at I/O. We shared the latest and greatest from Google AI Studio and Antigravity. As you can see, there's so many new updates and new features that we introduced today and more coming soon as well. So keep the feedback coming because that's how we can co-build these products together with the community. That's really important to us. I literally showed you how you can build a new business entirely from scratch with AI Studio in 20 minutes or so. And then you can take it to the next level with Antigravity. So, no better time than starting today. ANSHUL RAMACHANDRAN: And thank you. JOANA CARRASQUEIRA: Thank you so much. ANSHUL RAMACHANDRAN: Enjoy the rest of your I/O. JOANA CARRASQUEIRA: Thank you. [APPLAUSE] [MUSIC PLAYING]