Introducing Antigravity 2.0

本站添加: 2026-05-27

[00:00:00]Introducing Anti-gravity 2.0, a new standalone desktop app for orchestrating agents. Less than a year ago, we launched the Google Anti-gravity IDE, which included the agent manager, one of the world's first agent first guies on the market. Millions of users loved working in the agent manager, accelerating both their work as developers and non-developers. We've heard your feedback and built a bunch of new features. Combined with the recent leaps in Gemini model capabilities, we're excited to release anti-gravity 2.0 know and take a bold next step towards an agent first world. Let me show you how you can get started. So, let's go ahead and open up anti-gravity.

[00:00:35]We're going to start by onboarding.

[00:00:37]First, I'm going to sign in with my Google account. And if you're an anti-gravity IDE user, your next update will bring you exactly into this application. So, we'll go through onboarding.

[00:00:47]I'm a light mode kind of guy. Next, you'll see the build with Google page.

[00:00:51]This allows you to install skills and plugins that hook into your favorite Google products. So, in this case, we're going to install the modern web guidance, which helps your agent navigate the web and build the best possible web products. We're going to install the Chrome dev tools, and there's a bunch of other options that you can choose from. By the way, if you're already an anti-gravity IDE user, you'll see a screen like this, and you can go ahead and leave that keep anti-gravity IDE checked so that you can transition smoothly into anti-gravity 2.0. We'll start by creating a new project. Anti-gravity 2.0 is centered around this idea of a project. You can add folders, add git repositories, and this really serves as the launchpad for your agent. So, in this case, we're starting a brand new project. So, I'll go ahead and create a new folder, and we're going to be building a game. It's called Street Guesser. So, we're going to add that folder as a project and dive right in. So, first, I'm just going to ask it a quick question. Let's just warm up. What can you help me with? And it's a coding agent. It's also just a general assistant. It'll tell me some of the things that it can do, as well as some of the new commands and new features that are available in anti-gravity 2.0.

[00:01:50]One of my favorite features is using my voice to tell the agent what to do.

[00:01:54]Here, I'm going to press the mic, give it mic access, and tell it what kind of game we want to build. So, I'll press the button and say, "How would I build a video game web app that lets users put in a city name or draw a bounding box around a map, then let the user guess all the street names in that particular region?

[00:02:14]I'm open to using any API, but ideally keep it pretty cheap.

[00:02:19]And you can give me a couple of options on how this can be built. You don't actually need to build it yet, but I do want to see a pretty comprehensive plan for what my options here are. And why don't we add that we want this to be an artifact implementation plan. So, we'll go ahead and send that. And we're using Gemini 3.5 Flash, which you can see it's really, really fast at putting together a pretty comprehensive plan for how to build a game like this. We'll open up the implementation plan. You'll see there is now an architecture diagram of what we can expect. I noticed that Gemini had a couple of open questions for me to answer before it can actually put together a comprehensive plan. So, I'll go ahead and answer those questions just like I would in anti-gravity normally by leaving comments on the artifact. So, I'm pretty comfortable with the user allowing picking any part of the world. I really like its idea about making a classic paper map style game. And I think we're going to need a server. So, let's choose a full stack framework for this. And I'm going to ask it to update my plan just to be sure that I know exactly what the model is going to go off the build. All right, it's already gone ahead and updated my plan. So, I think I'm ready to start building. I'll go ahead and click proceed. Anti-gravity likes to keep you in control both from a safety perspective and it's just good for you to know when it's running commands on your computer. Now, of course, this is configurable, but in the beginning, I want to get a feel for what sorts of commands the agent is running for me.

[00:03:34]So, here it's asking me to accept, reject, and whitelist particular commands that it wants to run. In this case, I wanted to create this sort of app and I want to install these dependencies. So, I'm going to go ahead and press yes. But all of this is configurable, complete control over which types of commands you always want to run and maybe some commands that you want it to ask you every single time.

[00:03:53]You can go ahead and enable that inside of your settings. And we have a bunch of options, including things like sandboxing and network permissioning that can get your agent dialed to exactly what you want. And just like that, Gemini 3.5 Flash has already finished the first version of our game, and it's telling me exactly what it built via a walkthrough artifact. And now it's suggesting that it can actually run my project for me so that I can open it up. So, let's take it up on that suggestion. I'll say run it for me so I can check it out. The agent has now run npm rundev, which we can see right above our toolbar. And to view these sorts of background commands or background sub aents, we can go ahead and click it, and you'll see the output directly on the right. And now I can see that my server is running at localhost 3000. And by clicking on localhost 3000, I can open up Chrome and see a pretty cool, pretty nifty version of this classic map where we can go ahead and guess. So, let's go ahead and pick one of the coolest locations in San Francisco, North Beach.

[00:04:43]Let's go ahead and just try and play this game. It's kind of cheating because it did put a couple of the street names and the placeholders. So, we'll go ahead and start with those first. All right, I think those are all the streets that I can name for the purposes of this demo.

[00:04:54]So, we'll go ahead and surrender. And it looks like we got 14 out of 128. So, that was pretty cool for a first pass.

[00:04:59]Gemini 3.5 put together a very very good game here. But now comes the fun part. I have a bunch of ideas on various features and some of the design elements I want to add to my game. And this really is where anti-gravity 2.0 starts to shine. We can introduce parallel conversations and sub aents. And our agents are capable of a lot of sophisticated work. So here we'll start add a confetti animation. When you get a street right multiple times in a row, keep a streak counter and show confetti with increasing intensity the more you get it right in a row. The first one should not have confetti, but should light up the street counter. And here I'm going to go down and select instead new work tree. And what this will do is spin up a new folder with a copy of the code so the agent can run in parallel without interrupting my main flow. I'll go ahead and send that and I'll start a new conversation and we'll do something similar. Design me a trophy for when a user gets more than 50% of the street names correct. Use image generation to give me something that matches the app's aesthetic. So, while that's going, it looks like my first conversation has actually ready and giving me a plan. So, I'll go ahead and review this plan. It looks like it's going to add an external library. This generally looks good. So, I'll press proceed. And we'll start a third conversation because I have a lot of ideas of things that I actually want to build. This time, I'm going to work inside of my main conversation just so that I can work on this one a little bit more synchronously. I'll say instead of a search bar, show some presets on the homepage for a couple of popular regions. Let's do San Francisco, North Beach area, West Village, New York, Treasure Island in the SF Bay, and downtown Dallas. What's really nice about using a fast and intelligent model like Gemini 3.5 Flash, our agents are already waiting for us to give feedback.

[00:06:38]So, here it put together a pretty nice looking shield trophy for when someone gets more than 50% correct. And this looks pretty good, but I have some other ideas. I like this aesthetic. Let's actually have three tiers of trophies.

[00:06:51]We can do one for really really good achievement and then one for 50% and then one for 10%. Let's do three sub aents so that generating these is faster. We'll go ahead and send that. So it looks like adding presets to our homepage is complete and it's given me a walkthrough and I can also see all the files the agent changed. I can click on a particular file in this case app page.tsx and I can leave comments directly inside of this artifact pane.

[00:07:17]And I want to move this into a constants directory. There are some things that I want to change, but I can just leave comments just like any other artifact.

[00:07:24]And now I'll tell the agent what to do differently. So, while that's working, I have another one. It looks like the streak is done. So, this will be fun to check out. So, it looks like my image generation's finished. And I have a carousel of three different options.

[00:07:39]Those look pretty good to me. And it's actually modifying some of the code to make it so that these trophies pop up at the end of our game. So, I'll go ahead and reload my app. And as you can see, I now have presets for North Beach, West Village, Treasure Island, and Dallas.

[00:07:54]Going ahead and clicking around, things look pretty good. And why don't we go ahead and play one of these games? So, let's test my knowledge of West Village, New York. Perry Street, that's a thing.

[00:08:03]I do know that. Hudson, I guess we could do all the numbers.

[00:08:07]Starting with seventh. We'll try five.

[00:08:09]Four. Okay, four looks good. Three. And now we have a bunch of confetti coming on the screen. It's actually kind of fun. And we're on a streak of five. That is awesome. Okay, for the purposes of demo, let's go ahead and stop and try and evaluate what we should build next.

[00:08:21]So, we'll go ahead and press surrender and reveal. And it looks like our badges and trophies are showing up. So, I'm a novice explorer of West Village, New York. Great. So, the fundamentals of our game are now set up. But, I want to add some more complex features and really show off how anti-gravity 2.0 can really alleviate some of those gnarly technical problems. So, to start, we'll go back into voice mode and let's ask it to do something. Let's add a backend to this that connects to a SQL database. You can use SQLite for this web app. Have it store the user's favorite maps so they can access them later as well as their history of games and their scores for each and let them see their game history and see their progress over time and compete against their own high score. So this is a pretty complex task because it involves refactoring a bunch of things throughout our application. So you'll see it doing a lot of research both in the front end and in the back end to in order to put together a comprehensive plan. I'll go ahead and review this. It looks like it's putting together an entire database schema. It'll do some refactors on both the client and the server. And generally, this looks pretty good. So, let's go ahead and build it.

[00:09:22]But this time, I want to show you a new part of anti-gravity 2.0, which is sub agents. So, I'll ask it to spin up a bunch of sub aents to get this work done in parallel. Then, let me know when you're done. I'm excited to play this game when you're finished. And I actually had a second idea, which is why don't you test yourself? So, we'll use / browser. And we know a lot of people really loved seeing the agent use a browser in order to verify that it completed its task correctly. So I'm going to use / browser and ask it to test itself. So this is my first time setting up the browser. So I'm going to go ahead and go back to Chrome and follow these instructions. Go ahead and press remote debugging. Have it check again. And looks like we're good to go.

[00:10:00]Just like that. Couple clicks, our browser is now set up. Cool. So the agents are now working for me. And it kicked off three different sub aents. It has one to test end to end with the browser. It has another one that I guess is a UIUX designer. And then we have another database and API engineer. So roughly QA front end and back end. And it also added a timer for itself to check in on all the sub aent progress 90 seconds later. And if you head over to your artifact pane and click overview, you can see a list of all the sub aents and what they're doing. And if an agent gets stuck, you can easily see action buttons for you to oneclick approve. In this case, the browser wants to go to /dashboard. So, I'll go ahead and press approve there. This is one of my favorite features in anti-gravity. Not only can you manage many main agents, but your own agents can actually manage their own teams of sub agents. It's really common for some of my more technical, more complex engineering tasks for my agent to spin off five, 10, if not 20 different sub aents to go ahead and check each other's work or just paralyze the work so that it gets done faster and more correctly. So, with so many agents running, now is probably the time where I can go get a snack or grab a coffee and come back when things are done. Anti-gravity will notify me and let me know when my attention is required either to approve a command or just unblock some sub aents. All right, I'm back and it looks like our sub agents have finished and it's gone ahead and tested itself and verified that everything works correctly. So, let's go ahead and play this new version of our game that is now full stack. So, looks pretty familiar, I guess. Now we have looks like favorite maps that we can put in. Um, we'll go ahead and select a section. Let's try this part of San Francisco. And I can name it now. We'll call it North SF. And we'll go ahead and play North SF. 276 roads. Let's get started. We did this earlier, so we'll speed through a couple ones that we remember. Broadway, Market, Lumbard, Union, Stockton, Pulk.

[00:12:02]Great. 3%.

[00:12:05]And we see it's now a personal high score. Very cool. And I can go over to my archives and it's saving my progress just as we instructed it to. All right, let's just see if it works in other parts of the world like London or New York.

[00:12:20]So, wouldn't it be fun to actually use the agent to play the game that it just created? So, let's add /browser, can you play the game for me and get all the streets on Treasure Island, San Francisco? And I'll let it cheat. You could do web search research to make sure you get at least 90% accuracy. And we'll see how the agent does.

[00:12:40]So, here it's utilizing a sub agent to actually go ahead and play the game. And you can see how it's deciding it to get 90% or probably it will probably just get 100%. It's actually going ahead and querying an API to get a list of all the streets on Treasure Island. That's obviously not something that a human would do if they were playing this game, but it's an agent, so it can do whatever it wants.

[00:13:01]Looks like it's putting in a bunch of street names. Interestingly, none of them are registering. And let's see if it decides if this is a bug in the game or a bug in its logic. Looking at the thought traces, it seems like the model has figured out that it was not pressing the enter key after putting in each street name. And maybe that's why it wasn't able to get 100%. And just like that, now it's putting in all the names.

[00:13:22]Oh my gosh. So much confetti. 157. Oh, it actually missed one. Interestingly, it streak got reset. We're now at 30 out of 45. There we go. Personal high score.

[00:13:34]100% of the streets on Treasure Island done by an agent. Now, that's going to be a really hard score to beat, but it got that top score. Honorary decree, Grandmaster Cgrapher. Now that we've played our game, I want to set up something called a scheduled task. This is something brand new to Anti-gravity 2.0, and it lets you run prompts on a loop. So, I'll go ahead and start a new scheduled task. And I really actually enjoy this game, so I want it to tell me every morning a new random location in the world that I can play and maybe have it tie back to historical facts or something interesting about that particular region. And while I want it to be random, I don't want it to be a completely random part where there is no famous street that I just have absolutely no hope of guessing. So, I'll add in the prompt that it should be something with at least one or two famous streets. So, we'll go ahead and call this morning street guesser. We'll have it run at 5:00 a.m. every morning.

[00:14:24]And I'll put in my prompt. Choose a random big city in the world that uses English street names. Pick a sub region of that city with famous street names and give me a local host link to play my game every morning to start my day. give me a famous landmark or historical reference point for that particular bounding box. We'll add that scheduled task and within a couple of seconds, we'll now see it on our dashboard. Now, it is not currently 5:00 a.m., but I do want to show you what this would look like if it ran at 5:00 a.m. So, I'll go ahead and copy paste that prompt and just run it right now, and I'll show you exactly what that would look like if it was very, very early in the morning. But what's really cool is Anti-gravity 2.0 can run in the background even when all the windows are closed. So, I'll go ahead and X out this window. And as you can see on the top menu bar, there is a little icon. It's the anti-gravity icon that is telling me how many agents I have running and a quick button to reopen that interface. So if I reopen, I'll see that my agent has now finished.

[00:15:17]And just like that, the agent has decided to take me back to London. And I guess today I will play Street Guesser in London. And what's really cool here is that those scheduled tasks will continue to run in the background. And you can set up as many of those as you want. So, you might have a task that checks the status of a PR or puts together a daily brief of tasks that you have to get done today. The possibilities really are endless. And as long as you see that anti-gravity icon in your menu bar, you'll have confidence knowing that those tasks are executing behind the scenes and that your agents are working for you. And just like that, there's your first look at anti-gravity 2.0. You just saw sub agents working in parallel. You saw work trees, scheduled tasks, image generation, the list goes on and on. On behalf of the anti-gravity team, thank you for watching and we hope you love what we've built. Check it out at anti-gravity.google.