Google Antigravity: Hands on with our new agentic development platform

0:06Hi, everyone. Welcome back to Release Notes. My name is Logan Kilpatrick. I'm on the Google DeepMind team. Today, we're joined by Varun Mohan, who's the co-lead for Google Antigravity. Varun, thank you for being here. Let's dive in and talk about Antigravity. Do you want to sort of give us the TLDR and like the high level of what the product is and all that stuff? Yeah, so Antigravity is an agent development platform. Obviously, it has an IDE, but there are a couple sort of key differentiators we've come to market with to kind of start with. So in the very beginning, one of the things it has is it has a familiar editor that that all of us know how to use.

0:39But it also has this agent manager that allows you to orchestrate many, many agents that can operate over your code base. And in addition, a brand new surface, which is the browser. Agents now with the advent of Gemini 3 can also actuate the browser in a very, very capable way. On top of that, the agents are able to interact with the developer in these kind of indivisible sort of tasks that are in units called artifacts that sort of give the developer verifiable steps that the agent is kind of going through. and the user can kind of go in back and forth with these artifacts to kind of complete longer and longer tasks.

1:12Kind of the inspiration here is, and a lot of it was by seeing Gemini 3, we basically saw that models are able to now go longer and longer with less human intervention. And we want to take maximal advantage of this to kind of give the best possible experience to developers. Yeah, I love that. And I'm excited to deep dive on like all the nuance of anti-gravity and some of the innovation from a product perspective. Actually, just to sort of set the stage, you came over to DeepMind, joined Google to start focused on building this product, to do a bunch of agentic coding stuff, just to sort of like that experience of like coming over and like how it's been actually now seeing.

1:49And obviously I think you all were, you know, sort of training models in some capacity before, but like now being sort of super close to like the frontier foundation model perspective, I'm curious how that's like influenced your sort of like level of ambition, but like also excitement and like what you're thinking about from a product experience and like also just generally what has what has it been like to come and join DeepMind and do all this stuff? I think first what you said I think that resonated the most is the level of ambition. I think the level of ambition has increased tremendously and that's kind of because we we don't need to reinvent the wheel in a lot of different pieces right.

2:21The model is like an incredibly important piece to this experience right. We can we can see this if you have a bad model and it's not able to go and follow instructions properly not able to take many, many steps in a row, it doesn't really matter. You're kind of dead on arrival in a lot of ways. And I think one of the benefits of coming here that we've sort of seen is how much we're pushing the frontier of models, not only for coding, but across many different axes, like multimodal capability, UI control, all of these different pieces, I think are really, really critical to deliver a great experience.

2:54When I look at the kind of industry over time, you can kind of see how things have transformed, right? If you go to 2021, you see the advent of GitHub Copilot, right? And it had kind of basic autocomplete. And that was already kind of groundbreaking at the time, right? And then after that, you had some of these chat experiences with Gemini and ChatGBT. And that started to make its way into the IDE as well, right? A chat panel on the very side that kind of understood your code base. It was very personalized to all the private data inside a company or inside an organization. And then after that, you sort of had these agentic IDEs and agentic experiences.

3:27And I think the next version of this, it can no longer be the case that you understand code only. They look at the browser, they kind of do some research, right? They might read some papers. On top of that, they might look at their bug reports to see what tasks to ultimately solve. And I think that's what got me and the team very excited about the fact that our models are getting very capable, not only in the silo, but across many, many dimensions. This is one of the threads that I personally had not sort of deeply grokked until I started thinking about this, like coming up to these launches of just like how important, but also different and impressive it is to bring those like non sort of coding developer, even though developers are doing research and like non-traditional like developer tasks into this experience.

4:13And I'm curious, like how you sort of like think about finding the balance? Because obviously like developers do a whole lot of like they're on, you know, you know, meetings and they have calendar needs. And like, where do you sort of draw the line of like, what is like stops being a core part of that experience? And you would sort of like outsource out of that main product experience versus like, what do you pull it and like maybe research is one of those things and, you know, something like QA testing or deployments or stuff like that. Like where's that line in your head right now? Yeah. So when I think about what does a developer ultimately do, writing code is maybe a very small chunk of what they ultimately do, right?

4:47They write code, they review code, they test code, they do bug code, they deploy code, right? There's many, many aspects of this. And for each of those, they might actually have a completely different surface. I'll give you an example. Inside Google, we use Google Docs. And we might write up system designs inside Google Docs. And it's quite important, actually, to understand the system design if you're going to go make a large-scale change in the code, or you're going to review code, for that matter. And in that case, I think if all we did was looked at the code only, we'd be operating in a very narrow silo.

5:18We wouldn't be able to solve these very hard problems for the developer. And the way I like to look at it is it's almost like you're chopping wood, right? A developer has a hundred units of time. Some units of time is spent writing code. And we can, if we, all we do, let's say that's 20%. If all we do is we make it very easy. Once you know exactly what to do to write the code, I think there's a limit to how much more we can accelerate software development, right? Like, I think, I think basically if you look at the hierarchy of needs and what we're able to provide for developers, it's like, what, what should I build?

5:48How should I build it? And building it. And I think the models, if we're only able to do code, they help with the building it part. But how should I build it? It's actually a very, very interesting question that I think AI can help accelerate tremendously. And I think anti-gravity has a lot of the pieces to kind of push in that direction as well. Right. Who is actually the ideal user for anti-gravity? And both like, actually, there's lots of folks using it internally. It's available externally now. Like, who do you think is going to be like the most successful user persona today? I think it's very important in all of these products in general to be a dog fooder of the product, right?

6:23Otherwise, you're kind of just like implementing a bunch of features. And you kind of see these products where all they have is a bunch of features. And you're unsure exactly what's going to get used or what's not going to get used. So internally at Google, a lot of Google DeepMind actually does use the product. This gives us like a lot of rich feedback about how we can iterate even before we ship something to the public. For folks who don't have context, the internal Google code base, all this infrastructure is extremely, I would guess, on average, much more complex than other environments, like anywhere in the world.

6:53I think that's a really unique requirement and challenge of doing this actually inside of Google. It'd be easier to build something completely external that doesn't have to touch Google stuff. And I think the internal stuff is extremely difficult. Yeah, exactly. I think this is definitely hardening our product to work with the largest enterprises. I think if any enterprise kind of talks to us and is like, hey, is anti-gravity working on a very large code base? I think very few companies have a larger code base than Google. Surprisingly, yeah, because of the product and targeting all these developers, folks like Sergey use the product all the time.

7:30He's shipping a lot of CLs and PRs now, which I guess is a really good thing for him to be dogfooding the product as well. But I think at a high level, sort of, we're trying to make sure that a lot of the researchers inside the company are able to use the product. And the benefit here is they can kind of also see the capabilities of the Gemini models in real world scenarios. It also gives us a lot of rich feedback of what is it like for anti-gravity to be useful for solving very, very complex tasks, right? I think there's one thing for vibe coding very ephemeral applications. But I guess the real persona, and you got to it, is the developer, right?

8:05We want developers to be able to build their dreams, right? We want them to be able to build very, very complex apps that solve really hard business use cases. We'd like singular developers to build entire companies, right? I think this is like a line that's floating around a lot in the ether right now. But I think as a whole, kind of working with a lot of the researchers, we've also learned a lot, right, about the style of tasks that we need to be able to solve. And this dovetails back to the conversation of the model only being good at code. It's very clear if we want to accelerate a lot of research, which we are very excited to do, the model will need to do a lot of things beyond just write code, right?

8:42A researcher might fire off a training job. They might kick off an eval. They might need to analyze the eval. And this is like a lot of boilerplate work that they're doing, a lot of which doesn't sit inside the editor, right? And sits on many, many other surfaces. And this also was the inspiration for why we want the browser, right? It's not just for UI testing kind of simple web applications. it's for actually actuating the browser for very complex use cases that we'd like for the general developer. And how do you think about the, do you have a sense that this user persona will change over time?

9:17Like obviously the whole sort of like tailwind of like, who is the developer is changing, what it means to create software is changing. I do think the experience today like feels very developer centric. And do you think like over time you'll sort of like add abstraction layers on top of that? So one interesting thing we've already found is folks that are not technical are already using anti-gravity internally, which is very cool to see. I think the way I would like for us to think about it is we should be constantly trying to raise the ceiling of the product, but also making it really, really easy for folks that are just technical adjacent to be able to get value from the product.

9:53But I think my worry about us opening up the floodgates today is we kind of become a product that is not good for either persona, right? And that's like a big worry that I sort of have. Like in my mind, if I'm looking at a pure vibe coding use case where you want to quickly build an app and then have all the bells and whistles of AI infused in it, AI Studio is an awesome platform to go and do this. And I think we should not be trying to replicate that experience, right? We would love for people to use AI Studio for those experiences. And if they have an application that they'd like to kind of continue to iterate on over time or add more features to over time, we would like anti-gravity to be the right platform for that.

10:30This was, you and I were going back and forth on this. And I think it's actually an interesting, it's an interesting point to make mention of like, we will actually have the AI studio to anti-gravity path. We're still figuring out the details and making sure that it works well and all that stuff. But like, I do think like for a lot of folks, they start, even technical folks, they like start with this vibe coding journey and they're like, okay, great. I sort of can now start to materialize my idea. I am a software engineer, I am a developer, I really want to go deep on this and make sure I've made all the right technical decisions to scale up, et cetera, et cetera.

11:02And I think being able to like take things that you've vibe coded out of ASTudio, go to anti-gravity, I think is going to be this like magical experience. So I'm excited for that to land. Hope you're awesome. I think that's going to make these apps like more economically valuable to the people that are building them. A hundred percent. Yeah. I'm also curious like this, this sort of like broad industry trend of like agents versus the IDE. And obviously you sort of like have both of these. Do you think there's a world where like you stop, like the IDE actually goes away and you sort of are truly orchestrating agents?

11:36I think, and I don't know how much you want to talk about it, but like there's like literally like design product decisions that you all have been making the last few days about like, what is the default experience? Like what do we drop developers into? And I think there's actually, this is an interesting thread of like how do you meet developers where they are today versus like also showing them this future of what you could do. And yeah, I'm curious, like that tension actually, as you build this. Yeah. I think for the foreseeable future, the editor is not going away. And that's purely because I think developers need the control to ultimately see the underlying code and edit it right at the lowest level.

12:11Like taking that away is going to be like catastrophic in a lot of cases, right? But I think what ultimately does happen is maybe the time you spend actually writing singular lines of code in the editor, doing autocompletes and all these things may drastically go down over time, right? And then what ultimately ends up happening is the editor looks like any other work surface for a developer, right? If you think about it, there are many places that a developer generates work product. Editor is one, maybe inside their code review tool is another, maybe in a Google Docs or Docs or Sheets is another one.

12:40And I guess all of these become like important services that kind of an agent can operate on top of and collaborate with a human, right? But I don't think it is going to be the case that the editor itself is going to go away. I think the entry point of how much time people spend looking at individual lines of code in the editor, instead of looking at artifacts, which is a new kind of paradigm that we brought together, I think the trade-off will be artifacts will, the amount of time people spend looking at artifacts and verifiable sort of units from the agent is going to increase, right? And I think the nature of why that's going to be the case in general is let's imagine the AI is going to write most of your code and actually a lot of it is going to be the AI needs to figure out and the agent needs to figure out a way to make it easy for the developer to review the code without kind of kind of like making them write every piece themselves right so there does need to be this higher level abstraction that lives between the the kind of developer and the agent um and uh and yeah the editor may not be the the like the editor or the lines of code themselves may not be that interface.

13:40Yeah, I totally buy that. To be honest, I feel like it is this like evolution of like, what are, and like today you're reviewing PRs or CLs if you're internal to Google. And like in the future, maybe it literally is you're reviewing like the implementation plan or like the execution plan or like the final test suite or something like that. That brings me to another thread around sort of this like agent human collaboration and how like verification and autonomy and all that stuff is like baked into the product experience. Obviously today, like models can't run for at least, you know, reliably in a lot of cases aren't running for days on end or even like hours on end.

14:14It is a little bit sort of like faster of a feedback loop. So I'm curious how you all are thinking about in anti-gravity today. Like when is the human in the loop versus like what decisions can you just like fully delegate and let the model run with? So actually we give you a way in the product to kind of have this thing called agent assisted development, which is you can actually specified the level of autonomy you want to give the agent. And the agent itself can decide, hey, like for these kinds of terminal commands that it runs or for these tasks, does it want to notify the user? Does it think the threshold is high enough to notify the user?

14:46And obviously we have like evaluations internally to make sure that these are reasonable. But is this reasonable? Are these commands that are being run something that it should tell the user? And I think you do need to kind of trust that the AI will do this. But at the same time, you need to give kind of the escape hatch for the developer to kind of provide feedback. And that's why kind of in the artifact, you can asynchronously provide feedback anywhere. You can provide feedback almost like comments on a Google doc, right? On the artifact. Also on the chat panel, you can always write a comment and that will be folded into whatever the agent is doing.

15:19The agent will decide when the right time is to fold in this, right? So I think there's going to be a mix of everything because I see a world and I already see this right now at Google where folks are going and grabbing a coffee and coming back and they're looking at the work that the agent has done and they're like, hey, actually, I want to tweak it this way. And it will go and take that feedback and go and continue working for the next 30, 40 minutes, actually. It's kind of remarkable that this has happened in such a short period of time. Yeah, that is super interesting. I have a very tactical question, which is this, and we actually see this across the ecosystem for a lot of these coding agents today, is what's happening on server side versus what's happening locally.

15:56And I think this is an interesting distinction. And I think today, if my understanding is correct, is Edge of Gravity is running locally on my device. Obviously, it's connected to a server model that's hosted on a server somewhere. But, like, the code is being generated and stuff, and it's visible on my computer. How do you think about, like, that? And obviously, that means, like, I have to have my computer has to be on in order to continue to make progress. How do you think about that experience today versus, like, the server-side background agent sort of running in a container somewhere? And, like, the tradeoff between those two?

16:26and like do they converge at some point? Are they completely different use cases long-term as well? No, I think that they do converge. I think that's because this isn't, like you have to imagine a model or an agent that is able to go for a very short period of time and complete a task or can go a long period of time, can also complete the task in a short period of time. The model is not going to be somehow able to only think long but not think short, right? It would be a weird property. And I think that's what all of us here at Google and DeepMind are really pushing for, right? Models that can go for longer and longer and execute tasks and more and more complex tasks, right?

17:00And you're totally right. I think these tasks outliving the sort of developer's machine is going to be quite important. But I think all the principles that we just talked about, about being able to provide feedback and all these different pieces is very, very important. And I think the reason why it's important and why some of these asynchronous experiences haven't taken off nearly as much, at least in the market, I think is actually because I don't know about you, but I'm really, really bad about communicating my intentions to the model. Right. When I tell it, hey, do X in the back of my head, I have a hundred other requirements.

17:30Yep. And I'm not very good at specifying them. Like I could actually tell all of our users, hey, specify every requirement upfront, but some things you only learn after you see the code. Yeah. It's a very iterative journey. It's kind of like why you can make a PRD or, or a system design, but ultimately when rubber meets the road and you actually build this thing out, there's a, there's a gap between the two. You realize, Hey, there, this piece actually didn't fit in the way you thought it did. And because of that, actually, despite having a very asynchronous system, you need mechanisms to allow the user to kind of understand the progress so far and iterate on it.

18:03So I think all the pieces we have built into anti-gravity are going to really, really like help in that environment as well. I think you really want that there, right? I find it very hard to believe I have a very complex code base. I say do X and specify a sentence. It is going to come back with exactly the right thing I want, right? Because one approach that it could take is delete all my code and rewrite it all from scratch and then do X, but I wouldn't really want this. And I actually think this dovetails into one last thing that we do have in anti-gravity that I think helps with this experience, which is that I think one of the things that I love about Google as a product as a whole, and my goal is not to just say only positive things about Google, is it almost feels like it does read your mind and understand context.

18:42Like for instance, if I just sort of searched a lot about a basketball game and then I wrote Curry, It's gonna show me Steph Curry. It's not gonna show me chicken curry, right? And I think you want the same experience also when you're using the Asian. And because of that actually, self-improvement is kind of built as a key sort of primitive into anti-gravity actually. So as your - This is the knowledge panel. This is the knowledge panel. So the idea here is like, as you're using the product, it is actually building up knowledge of your usage. So the next time you use it, it is not re-deriving everything, right?

19:14Because you can imagine these conversations that you have with the Asian, They're very rich in information. They're not only, they teach the agent things that are outside of the code base. Because a lot of knowledge, it's kind of like what you said, right? There's a lot of meetings that we do have and that builds intrinsic knowledge that doesn't live inside the code base. And the code base might be a little bit out of date on what the actual intent of the user is. And I think that kind of helps you bridge the gap. But despite all of that, I don't think that this is enough. I do think you want an interactive experience, even if the model is able to go out for days, weeks, and months, right?

19:44Imagine the idea of someone telling you, hey, Logan, go implement this feature. And you're like, okay, the agent says it's going to take three weeks. You look at it at the end of three weeks and it's wrong. You're going to be like, it's bad, right? Or at least they need to keep getting bumped at that point. Yeah. And there's so many human analogies where if you were an engineering team, the worst outcome is I, as a product person, say I need X, Y, and Z feature. And then I show back up in three weeks after they've completely done everything. And then I'm like, this actually isn't what I want. Like you check in, there's an internet process.

20:15So I think it like mirrors what humans do today as a best practice. As you have been talking, and I think more broadly in this moment of like, you know, the anti-gravity product coming together and sort of agents coming together, it is interesting. One of my reactions is just like how, or the questions in my mind is just like, how much do some of these like agentic coding experiences generalized across domains that have nothing to do with coding? Like as you're describing like, you know, the human in the loop components, I want to be able to like have knowledge. I want to be able to like do tasks in the browser and research.

20:48Like I start to imagine I'm like, this is just like all like all work fits in. And obviously some work you're generating code for some work you're creating other artifacts. And I'm curious, like, yeah, maybe there are a bunch of bespoke things that will continue just for software. But have you thought much about like how much this generalizes to just an agent overall? I guess the hard part about this is this goes to show the power of these models. Yeah. Which is that they are generally very intelligent. Yeah. Right? And that enables you to solve many tasks. And I think you're totally right. I found myself using anti-gravity for very different things.

21:22Right? I think recently I asked it to analyze a bunch of metric data. Right? And this metric data had nothing to do with the code or anything. Right? Actually, it was on some website. I wanted to analyze some sports stats. Yeah. And actually wrote a very local script. It did some stuff. generated me an artifact, a plot. I iterated on the plot a little bit in image space. And then after that, I got a brand new plot. And now I've been kind of like using this internally, just to kind of like analyze. I guess maybe to give you an example, I just started getting into baseball and Shohei Otani seems like kind of otherworldly figures.

21:54I've been just analyzing other baseball players to see, is there anyone remotely close? Surprise, surprise, no one is remotely close. But I just wanted to kind of get some metrics on that. And I think it's very, very easy to see how this generalizes beyond just uh beyond just software development i guess the the reason why we're not talking about this is if it's a very general purpose tool so i always have kind of a deep worry that hey if we go and try to make the product great for everyone we will make the product great for no one yeah right and that's like a deep worry that i sort of have and i think we deeply understand the developer we're all developers um ultimately and i think i think one One of the things that is true, that I think is exciting, is the number of developers in the world has been growing.

22:40So maybe that's the right way to kind of formulate this, is the number of people that are builders and developers is going up. And that's just because the ease of building has been going up. You look at the very beginning, people were writing assembly. Yeah. Very, very hard to. Horrible. Yeah. Assembly's horrible. If you haven't written assembly, don't write assembly. Yeah, exactly. And then after that, you had C, right? And then you had C++, Java, Python. And Python, if you just look at the gap in what a Python developer needs to know and what an assembly developer needs to know, it almost looks like they are not the same thing.