Google Cloud Live: Getting started with Antigravity

[MUSIC PLAYING] [BACKGROUND CHATTER] [UPBEAT MUSIC] SPEAKER 1: We are looking at a 5,000 share order.

STEPHANIE WONG: Hello, everyone, and welcome back to Google Cloud Live I'm Stephanie Wong, and if you tuned in to our last few episodes, we spent a lot of time in the terminal with Gemini CLI looking at how AI helps us write code faster.

But we all know that writing the code is only half the battle.

The other half, and let's be honest, usually the heavier half is getting that code to run somewhere other than your laptop.

So today, we're talking about Antigravity.

It's a tool designed to literally lift the weight of infrastructure management so you can focus entirely on building.

And so to walk us through it, I have two of the engineers actually building the engine.

So please welcome Kevin Hou and Andy Zhang.

Hey guys.

KEVIN HOU: Hello.

ANDY ZHANG: Hello.

STEPHANIE WONG: Welcome back to the show, Kevin.

KEVIN HOU: Thank you.

Always good to be back.

STEPHANIE WONG: Yeah, and welcome to the show, Andy.

ANDY ZHANG: Thank you.

Good to be here.

STEPHANIE WONG: Yeah, so you both have been working on Antigravity since day one here.

How has it been so far?

KEVIN HOU: It's been fun.

We've been doing DevTools for some time now.

Antigravity is probably one of our proudest, biggest achievements.

ANDY ZHANG: Yeah.

KEVIN HOU: And we're really excited to see it out in the world, and seeing so many people using it, and all the different use cases that have been fallen out of it.

So it's been really a magical experience for us.

STEPHANIE WONG: Yeah, how about you, Andy?

ANDY ZHANG: It's been so much fun working on Antigravity.

As an engineer by trade, I love building, and I think it's a really big privilege to also be building a tool for building, especially building Antigravity in Antigravity.

And I think especially now that it's out in the wild for several months now, it's been really amazing to get a ton of feedback from users and see how exactly can we improve the product to make it the best it can be.

STEPHANIE WONG: Yeah, it must be special to work on something that you're like, I use this.

I am an engineer by trade.

And I'm building something that I can see myself fundamentally changing my workflow myself.

So it'll be great to hear how you're actually using it in your personal workflows.

We do have a lot of people, by the way, tuning in on the live stream, everybody on YouTube.

We have people on LinkedIn.

A lot of people are calling out your videos, Kevin, and saying, hey, video on Antigravity is fire.

KEVIN HOU: Good.

STEPHANIE WONG: Thanks, Kevin.

And yeah, just a lot of people coming in from all over the world, so thanks, everyone, for tuning into the live stream.

As always, we'll be back week over week and probably talking more about Antigravity throughout the year.

OK, so moving on just to dive into the questions a little bit, before we get into the demos.

So I have to ask.

You both are engineers.

You have felt the pain of both building and deploying code.

So what are the specific friction points or nightmare scenarios in the past where you were like, I really wish we had an agentic development tool that could solve this?

KEVIN HOU: Nightmare scenarios, man.

STEPHANIE WONG: I don't know about nightmare, but friction.

KEVIN HOU: I mean, we've all had our fair share of nightmare scenarios.

I think trying to avoid those nightmare scenarios in the beginning is probably how we arrived at Antigravity, or just some of the DevTools that we've been working on, I think specifically, I mean, if we rewind three years ago, we were building autocomplete and using autocomplete products, and it was a completely different world back then.

And so I think over time, we've been more and more engrossed in this magical world of LLMs. And AI is doing more for developers.

And I think we saw a really special opportunity about eight months ago or so when we started working on Antigravity, where the LLMs and agents got so good that we had that "light-bulb" moment where it was like, oh my gosh, we could actually build a product that is different from anything that developers had used before.

Arguably, you could look at the code less and less and delegate more and more to the agent.

And that was one of the magical experiences where we're like, all right, we need to build a new product.

And that product was Antigravity.

STEPHANIE WONG: Yeah.

ANDY ZHANG: And as we saw how models got increasingly better, the thought that came to mind is typically when you're modifying code, you're living inside the IDE, and we had the side panel reserved for the agent.

And increasingly, we were wondering if most of our time is spent on the side panel and the agent inside of it is just more and more powerful, how can we best leverage that?

It's like when you're imagining have an entire table of space, and if you're only working in one corner of it, are you underutilizing the entire table.

And so I think the thing that made us zoom out and think back is like, hey, what would it look if we really made this a first-class experience?

And so alongside or within Antigravity we also launched the Agent Manager, which we're going to be doing a demo of shortly after, talking about how we could create this first-class agent user experience.

STEPHANIE WONG: Yeah, yeah.

We're excited to see that in action.

And also talk about how ways that people can use it beyond what they typically would interact with these types of tools, even if you aren't a traditional software developer.

So for those who just are learning about this for the first time, I guess just to dive into a little bit more about your engineering background.

Before you built Antigravity, what was your guilty pleasure way of building and deploying code?

Was it dragging and dropping files, via FTP?

Was it SSH-ing into a server and editing code with Vim?

I think we've all been there.

KEVIN HOU: So for me, I love using Vim and I did not like having to manage many different services.

So I was thinking, what is the best way to bundle all of those?

I would typically deploy things to GitHub pages because it was a two in one, one as the source control, and number two is it can also deploy my static pages so I can see changes locally out in the web, and share a link with any of my friends immediately.

STEPHANIE WONG: How about you?

ANDY ZHANG: Mine?

I will actually change it from what I said before.

Originally, I was going to go with GCS, Just kind of force push to whatever bucket.

But when you said SSH, it actually made me remember I was a big like Raspberry Pi, exposed my IP to the public internet, and just SSH-ed into that thing.

It had 1 gig of RAM and you'd just shove files onto that thing and it would go down every day basically.

But, it worked.

It was fun.

STEPHANIE WONG: Yeah, I mean, hey, you can tinker and you have a Raspberry Pi, why not?

ANDY ZHANG: It was fun, yeah.

STEPHANIE WONG: Amazing.

So last week, we were talking a lot about Gemini CLI to write code, and now we're looking at Antigravity.

So I think some people might be thinking, is Antigravity just Gemini for Ops or is it actually a fundamentally different engine under the hood, and how do these two worlds of AI code assistance and infrastructure platforms actually come together?

KEVIN HOU: Yeah, so the way we think about designing Antigravity is we wanted to make it as approachable as possible, so we don't want to narrow it on specifically one use case.

And obviously, I think Antigravity is amazing at many use cases thanks to the models that power it.

But ultimately, when we think about it, it's great for coding, but it's also really good for research and understanding.

It's great for iterating on things such as your travel plans for the summer.

Or it's really good as an education tool to learn about new concepts.

So for me personally, I've found I've started relying on Antigravity not only just for coding and building things like Antigravity, but also trying to learn new concepts that can help with my work.

STEPHANIE WONG: Yeah, exactly.

I think some people might also be thinking like, OK, is this just a platform-as-a-service tool?

Is it a dashboard?

So how would you characterize Antigravity if you could do it in one sentence?

KEVIN HOU: That's going to be a hard one.

I think that if I had to characterize Antigravity in one sentence, I would think of it as a really good tool for thinking, because typically the best thinking leads to the best work.

And I think that historically, I feel like a lot of tools have been more focused on the building.

And I think that Antigravity has a little more emphasis, thanks to agents, on the thinking aspect, which leads to the great foundation for building something great as well.

STEPHANIE WONG: Yeah.

ANDY ZHANG: I'll add in a couple more sentences because I couldn't squeeze it all into one.

I think the product has been built to enable all sorts of builders to come to the platform.

And I think one of the magical parts about it is we started with the software engineering persona, and thinking about that user as someone who was willing to grapple with the ambiguity of an agent.

Oftentimes, you have to provide guidance or nudge it in the right direction.

But starting with software engineers was a great way both for us as software engineers ourselves to think about the product, but then also a set of users that was willing to be early adopters.

And I think we're starting to see that group grow larger and larger to include folks that are technically adjacent, people that might not have ever opened a coding editor, or they're now approaching Antigravity or they're able to approach Antigravity and do tasks that they otherwise would never have dreamed that a coding agent specifically would have been able to do.

But I think the beauty of this is Antigravity and the platform that it sits on, and Gemini specifically, is a very good general use, multimodal model, and it's able to do a lot more than just literally write code.

STEPHANIE WONG: Yeah, and I actually position myself, I think, in that bucket of users.

And I was just actually watching a three-hour course that a YouTuber made about Antigravity.

And it's incredible the creativity that people have to build things.

It's not just, hey, I need to build a website.

But they're taking inspiration from other sites, and they're inputting that as an image file, and they're creating all sorts of ways to include their brand assets into it.

So it's really cool to see that.

We're going to dive into an example, actually, of how to build in Antigravity and be creative and think.

But how can people get started?

Do you have any advice for just how to be successful on Antigravity?

What should people watch out for?

KEVIN HOU: Yeah, so Antigravity has been used in a variety of different ways.

We have a lot of people that are very, very passionate about our product.

And to that, hopefully you're watching.

Thank you for that.

But we also want to just make sure that you're following the guidelines of how to use Antigravity, specifically with regard to third-party proxies.

We've become overwhelmed, almost a problem with success, overwhelmed with the amount of demand for the product.

And we want to make sure that people are using it in the right way.

And in order to give the best experience possible, just make sure you stay within those guidelines, those terms of service.

We want to deliver the best product for you.

And we believe that Antigravity will be that product for you.

STEPHANIE WONG: Amazing.

Well, let's just dive into it.

I feel like we can see exactly how to build within those guidelines, and see what you can actually do successfully on the platform.

So, Andy, I guess I'll pass it off to you to show us the Agent Manager and Artifacts and what all this means.

ANDY ZHANG: Yeah, let's dive right into it.

So right now, I have the editor open within Antigravity.

So from here, you can hop into the Agent Manager, which is the service I talked about earlier, where you can have a first-class agent experience over here.

And then from here, I'll just go ahead and open a workspace.

So here, I already set up an OlympicMedalApp over here.

Now from here, what I'd like to do is get a better understanding of what's been going on with the Winter Olympics recently.

I think there have been so many cool, iconic moments, like seeing Alyssa Liu's crazy performance and-- KEVIN HOU: [INAUDIBLE] ANDY ZHANG: Yeah, seriously.

STEPHANIE WONG: Yes.

ANDY ZHANG: And it's been awesome to see, but I feel like I haven't gotten a deeper scoop.

And so I've loved using Antigravity as a way to research and get a better understanding and catch up to things.

So I figured we can use the Agent Manager alongside Gemini 3 Flash to get a better sense of the medal breakdown and maybe go from there.

So I'll go ahead and ask, can you explain to me the mental breakdown for the Winter Olympics 2026?

So it's just a super simple prompt asking the agent if it can explain to me what's been going on.

I think it's top of mind, and I wish I had a lot of time to keep up to date.

But luckily, I think Antigravity is able to do a lot of the research for me, summarize things, and hopefully create an artifact explaining to me what's been going on in the Winter Olympics.

KEVIN HOU: Maybe it's a good time to just explain the artifact concept, and just generally what you're seeing on the screen, especially if you haven't used Antigravity before.

So this is our agent.

This is the Antigravity agent.

And as you can see, there's what we call tools or steps.

And within that, you saw Andy's first prompt which was all right, the instruction.

Purely five, six words or whatever, 10 words or whatever.

And then all of a sudden it's going to do all this work to accomplish that goal.

And including, if you scroll up that one about searching the web, and now it's putting together what we call an artifact, which is anything that the agent puts together that helps communicate information or intent to the user.

So I guess we can see that here now.

ANDY ZHANG: Yeah, so in this case, it created this artifact of an implementation plan.

And before we even dive into an implementation plan, I'm a bit more curious about the actual medals themselves.

I don't want to interact with the website just yet.

So can you show me a table of the medal breakdown per country first in the plan?

And so the great thing here is it's definitely a collaborative process.

I think of the agent as someone that I can talk with, I can provide feedback to.

And so in this case, I was hoping for some specific information.

And sometimes we would make assumptions about what I'd want.

And so that's the beauty of it being a conversation is I can clarify things that it might be missing, just so that I can iterate and get to the desired outcome.

So in this case, I'll just ask it if you can include a table.

And as you can see here, within just a few seconds, it went ahead and created a table of the Winter Olympics medals.

So really cool and great for Norway to see that they have the most medals this year as well as the most golds.

I am kind of curious, what is the a medal breakdown per gender.

So I'll go ahead and ask, break this down by gender here.

And so the great thing here is that I highlighted a specific part of the table, and I added a comment.

And if I sent this comment over here, the agent wouldn't exactly know what I was looking at.

And so while you can share context through words, similar to how you're talking to a colleague or a friend, a lot of the things are unspoken.

The actions speak louder than the words.

And so in the same thing here, when you're interacting with the agent and its work, a lot of it is implied through the actions and not the words themselves.

And so in this case, I was specifying, oh, like in this table, can you break this down by gender.

So I'll go ahead and send this comment over and then see the agent iterate on this, do some more research and update the table.

KEVIN HOU: So what do you expect to happen here with the implementation plan?

What are we going for?

ANDY ZHANG: Yeah, so what I would expect here is I asked the agent if they can break it down by gender.

So it's going to go ahead, grab the relevant data, so in this case, you can see-- KEVIN HOU: A lot of web searches.

ANDY ZHANG: --a lot of web searches.

It's working in parallel.

So you know it's working really hard for you.

And so what I would expect is it grabs all this data and it figures out how do I blend this data into the table.

And I think this is one of the powerful things about agents that is not talked about enough is that agents are so good at synthesizing information.

That was literally what they were trained on, where they just grabbed a lot of information and figured out how to distill it into this thing that can share any kind of information you're looking for within seconds.

So as you can see now, it went ahead and did a fantastic job of breaking down not only the gender, but also the total, did the math for me.

And so now I'm able to visualize all the top five, and maybe we'll go ahead and do a bit more iteration.

So what I am curious about is also how does this compare to the past.

And I think that it would be great to see how does Norway do compared to 2022 Olympics.

So let's see, compare to 2022.

Here, so I'll go ahead add another comment.

KEVIN HOU: I think one of the inspirations for this feature was, A, how do you interact with people in general.

They wouldn't necessarily just brain dump their chain of thought, which is essentially what you see in conversations these days, and really treating it as, all right, here's a document that you're going to hand to your peer, and that's what they're going to review.

And in many ways, this is a GitHub or Google Docs style of commenting where you can leave a batch of comments, review it at the top, and then iterate with another user in that way.

So we tried to model it after this.

And this is one of the most popular aspects of the Antigravity product, to be able to iterate on a document as opposed to reading this 1,000-word often more stream of consciousness from the agent.

STEPHANIE WONG: Yeah, yeah.

As you were demoing, I was thinking short of having a physical colleague sitting next to you that you're collaborating with, this is getting close to how you would explain, hey, looking at this section, I think we need to iterate here.

I would actually to add gender comparisons here.

So I feel like we're getting as close to collaborating with someone remotely essentially, and giving it eyes, and a sense of what's going on.

KEVIN HOU: Totally.

STEPHANIE WONG: So, it's great.

ANDY ZHANG: And it's great that I'm collaborating with someone who can get results so fast.

It finishes so fast here.

And the nice thing about it being really fast is that it allows you to have so much more iteration cycles.

And the more iteration cycles you have, the faster you're able to get to the vision that you have here.

So zooming back in, it created a 2026 versus 2022 comparison.

It also has a gender breakdown.

And so at this point in time, depending on what hat I'm wearing-- right now, I'm feeling like I just want to see what the medal count was, and then also compare to how it did historically.

It seems like Norway got more medals than a few years ago, and so did the US, as well as Italy.

Let's see if there's any other important information here.

Let's see.

Cool, this is an interesting, fun fact.

Maybe there are some other fun facts that we wanted to do.

Are there any other interesting highlights or tips that happened?

So I'll just go ahead and send this in the background as I continue reading this.

And the great thing here is I can also parallelize my work.

While the agent is doing some work, I can also continue reviewing.

So right below here, below all of the medal counts, it also outlines a plan for how it's going to build this.

In my case over here, I already trust the agent.

I know it's going to do a fantastic job at that.

So I can happily gloss over it, or if I want, I could just tell it, no need to include this, just to keep the plan more succinct.

And the great thing down here is it also shares a verification plan.

And this is something that we'll talk a bit more afterwards, which is how exactly Antigravity thinks about testing your work and making sure that you can trust the work.

Because it's not just about doing the work, it's also about verifying and making sure that it has proof that the work was done properly.

KEVIN HOU: Why don't we try building this?

ANDY ZHANG: Yeah.

Let's go ahead and click on the Proceed button.

KEVIN HOU: So yeah, the nice part that we called out, you could work in parallel, and I think the Agent Manager was designed in this way.

You can have multiple conversations going at once.

You could have multiple different workspaces up.

A workspace is like a project.

You can have multiple projects going at once.

But we're also using Gemini Flash, which is an amazingly fast and still amazingly powerful model, especially for coding.

And so this has been often on our team at least the go-to.

When we're building Antigravity, a lot of folks like to use Gemini Flash because it just lets you stay in flow state a bit more than just waiting for a conversation to wrap up.

STEPHANIE WONG: Right, so I mean, if it is a long-running test, you can walk away, you grab a coffee, or you can do multiple tests at once.

KEVIN HOU: Or you can do multiple tasks.

STEPHANIE WONG: Have 12 projects-- KEVIN HOU: So you really have no excuse anymore.

ANDY ZHANG: There's always something to do.

KEVIN HOU: Always something to do.

STEPHANIE WONG: Yeah, exactly, next-level productivity.

ANDY ZHANG: Cool.

So let's see what the progress of it so far, so as you can see, currently it's trying to implement it.

And it seems like it's going to start trying to test it out.

So it went ahead, opened the browser, and the really great thing here is that it sees the fact that the website currently cannot be reached.

So if we go back, it's going to see, OK, clearly I might be doing something wrong.

So how can I correct my own work?

And it's going to continue iterating on this until it is able to find a solution.

Or I can always jump in and try to help course correct its work.

KEVIN HOU: Looks like it wants input.

This is another one of these things.

Obviously, the browser is an immensely powerful service, and we can talk a bit more about why we built the browser, what it can do.

But an important aspect of this is that a browser can access anything.

And so we want to build in the right safeguards to make sure that if you're going to a particular URL, the agent literally can click and scroll on the screen, which is immensely, immensely helpful.

But we also want to make sure that the agent is operating within the bounds of what we expect, which is why we have approvals and things like that.

The way that I use this often is I go half and half screen, I'll have the browser up on the left, and then I'll have the agent manager on the right.

And it's kind of fun to watch, and maybe we can try and resize to see how the browser is doing.

But you can see that there's something up there.

And I was hoping that we could get to see it actually actuate.

Actuating is when it has that blue dot, and it's moving around and clicking and scrolling.

But it looks like it already finished up its work and put together the screen recording.

So it was a little bit too fast for us.

Maybe we'll catch it on the next pass.

ANDY ZHANG: Yeah.

Speaking about the screen recording, I think one thing that we can see here is that you can see a screenshot of this.

And you know how I was showing you guys that you can select text and you can comment on it?

One of the things that actually Kevin built, which I think is an amazing feature, is you can be even more specific with the context that you provide to it.

So for example, in this case, it provided here, I can go ahead and click on this image and then I can select on things that I want to comment on.

So maybe let's say I don't like this header.

Can you make it super blue instead?

KEVIN HOU: No, please?

ANDY ZHANG: Please.

KEVIN HOU: There you go.

ANDY ZHANG: There you go.

Maybe add a bit of emphasis, so it actually listens to me.

Cool, so I'll go ahead and fire this off and let it parallelize.

STEPHANIE WONG: I was going to ask if you could draw over images or sections because highlighting is great, but if it is an image, I think dragging is even more useful.

KEVIN HOU: Yeah, I love the image.

We also have image gen in the product, which I think also inspired this.

So now suddenly you've got website screenshots, you've got recordings, you've got Nano Banana, GemPix stuff inside of the agent.

And so we wanted the ability to similar to highlighting text, we wanted the ability to annotate an image.

And so now we can see, all right, we've got that blue border.

Let's see what it wants to do.

It's going to start scrolling and start screenshotting.

And it produces a playback.

So it's a very, very neat feature.

And I think also more importantly, if we zoom out and think, all right, what are we doing here?

We are starting to verify the results of agents much more than the actual how did we get there.

Obviously, how we got there is important, and that's why you can dive into the conversation if you want to.

But if you're firing off and this is our goal and our dream, and I hope that every user watching this will give this a shot.

If you're running three or four agents at once, you're adding a button here, you're adding a new page there, you don't really want to or care to look at all the individual things that it did.

You really just care to look at, does it work.

Does it look good?

And that's where the verification comes in.

So it'll screenshot itself.

And much like reviewing a PR, for example, this is the way that we see interacting with agents and how you can scale your mind space across three or four different changes that might be happening at once.

STEPHANIE WONG: Yeah, that's fantastic.

We actually have a lot of excitement, I think, about the browser extension as well, which I know you haven't talked about just yet, but you could actually play with this in the browser-- KEVIN HOU: Totally.

STEPHANIE WONG: --directly.

So we have a question actually from Dev, Ryan here wants to ask, so I actually find the Antigravity browser integration to be leagues better than some of the other similar tools out there.

So what's going on under the hood to actually power this?

Is it a special model training?

What is the extension actually do?

KEVIN HOU: Sure.

A lot of layers to this, thanks for the question and thanks for the complement.

That means a lot.

So we worked really hard on the browser.

And the browser is an example of there's a lot of things that are cutting edge about Antigravity.

One of the more unique ones is probably this browser integration.

I think you touch on a number of things here, Ryan.

The first of which is how is this powered.

It's actually an orchestration of multiple agents, and this is one of the first times we've seen different types of models playing very well together.

So for example, when you look at the conversation, there's a main agent.

In this case, it's the Flash model that is deciding what to do, delegating tasks, and it will choose to delegate what is called a browser task to what we call a sub-agent.

And this is basically given a goal.

This other agent will go off and purely try and accomplish that goal.

So in this case, I don't know if you can see it on the screen, but it probably said something verify that this page or this table loaded correctly.

And that's the only job of that agent.

And that agent will actually use a specialized model.

It's actually one that everyone can have access to publicly, but it's the Computer Use model.

And computer use is defined as a model that is capable of screenshotting and essentially deciding what to do next.

So in this case, it probably said something, verify that the table exists, scroll, and click buttons if you need to.

And it will take a picture of the screen and decide to scroll a bit, and maybe zoom in on a portion of the screen, and use that special capability, and then report back to that main model.

So it is this orchestration of multiple things happening at once.

This is a hugely powerful, complex part of the Gemini training world as well.

So it really is a coming together of a lot of the multimodal features that you see with Gemini.

Some of it is very specific to the way that we trained, specifically the Computer Use model.

And it's really designed to just actuate on top of a browser.

STEPHANIE WONG: Yeah, yeah, I think it's fascinating because folks might not know that it is this integration between specialized model and passing it back to the main model that's incorporating the feedback that it's collecting via the browser integration.

We talked before this session actually about how you can use that to your advantage for testing your applications.

So how does that actually really affect how people can move quickly?

KEVIN HOU: Yeah, I have my own ways that I use the browser.

I'm happy to hear you chime in as well.

I've used it.

So on one hand, there are front-end changes where a local host will spin up.

It'll click through the pages that I made, and then I actually have some instructions in my own prompts to just say, hey, validate that some of the other pages worked.

And in this way, it will not currently replace integration tests and the usual sort of things, but it gets a finger in the wind of, all right, none of the other pages are broken in any particularly glaring way.

The other interesting way that I use the browser is I actually have it authenticated on a number of things, like Google Docs, for example.

And I will use it as a primary browser to scroll through design docs, to look at public documentation.

And because it's almost like the agent is looking over my shoulder because we have access to the page content, I can just say, hey, implement this.

And it will be smart enough to know, hey, this actually means the doc that I'm looking at.

And in this way, you bring all that richness context into the agent's execution.

And that's my favorite way of using it or one of those pro-user hacks.

STEPHANIE WONG: Right, so you don't have to copy and paste everything.

You refer to it.

And then for testing, if you're an independent developer, it's like of one.

You can have some level of UAT early as you're iterating so that you just move on.

KEVIN HOU: Exactly, exactly.

ANDY ZHANG: I think I'm definitely not as much of a power user as Kevin.

I think a lot of how I think about it is, what would I do.

And can I just help the agent to do that?

And that's a really good litmus test, is like, if I'm doing this repetitively, what would it take for me to tell the agent to just do it for me?

So if I'm constantly testing this one flow, I'm like, OK, instead of doing this thing that takes two minutes 10 times, what if I just type it out for two minutes and then save 18 minutes of my time?

And that way, throughout every one of those iterations, I decrease how much time it takes for me to iterate on it.

And I think that's my favorite way to be using the browser, because at the end of the day, I have a vision in mind of how I want it to look, and I want to be able to iterate as much as I can to it.

And I know exactly what is the step-by-step flow of what it should be testing.

And so I can clearly outline to the agent like, hey, this is what I want you to test out.

And then go ahead and test it every time I ask you to make some changes.

STEPHANIE WONG: Yeah, and I do know there are even more features, like for any repetitive tasks that we probably won't get into today.

But I did learn about skills and how you can implement.

ANDY ZHANG: You can do a whole episode on skills.

STEPHANIE WONG: Exactly, we'll do that next time.

But anywho, that's great.

Anything else you want to show with the Olympics use case here?

ANDY ZHANG: Yeah, maybe just some final touches here, I think one cool idea is how can we visualize this, not just in a table format, but if we can create a heat map of all the countries and the medals that they have.

So can you make a heat map of all of the countries based on the number of medals that they have?

So this could be one last final touch that we'll do.

Right now, it's just a usual table that you can probably find searching through Google.

But what I think would be really interesting is different ways of visualizing that data.

I think it's a cool reminder to see all these different countries around this world and seeing which metals are where.

So let's go ahead and ask the agent to create that heat map.

And once again, it does a lot of the research.

It analyzes the plans.

And then from there, it figures out how exactly it wants to build it.

STEPHANIE WONG: OK, cool.

ANDY ZHANG: Cool, So it's hard at work right now.

It's updating all the files it made before, as well as making new ones.

Let's see.

So you'll notice we're spending a lot of time in the Agent Manager as opposed to the editor.

And I think that's one of the high-level-- for folks who are more tapped into the industry, this is in many ways, the direction things are heading.

These agents have gotten so good that you can just trust Gemini will, of course, verify that it worked correctly, but you spend less time looking at the code.

Of course, things like autocomplete and looking at files, that's still an amazing experience inside of the Antigravity editor.

But for a lot of tasks, I think we find and our team finds that we end up operating at this higher level.

If you're putting together a design doc, a very common thing you want to hand off a feature to another team, you'll often spend some time in the Agent Manager just putting together research and maybe asking and iterating with the agent.

Put together this artifact to describe how this change is going to go.

Share that out with your colleagues and use that as the starting point, as opposed to what would traditionally have been sitting inside of the editor and clicking through a bunch of different files.

It helps you ingest all that context, all that information a bit easier.

STEPHANIE WONG: Yeah, it's a bit more intuitive and actually related to that, we actually did get a question from the audience ahead of time from Dev as well.

And heckno is asking, will the editor evolve to look less like a traditional editor, as AI-assisted editors become more of the norm?

KEVIN HOU: Yeah, I guess there's a couple of definitions here, traditional editor, AI-assisted editors.

We actually took maybe even a more extreme stance and said the Agent Manager, we wouldn't even classify as an editor.

There's no way for you to make a text edit inside of the Agent Manager.

And this is by design.

I think we want to give people the AI-assisted editor, the escape hatch, the advanced I need a double click and go into a particular file.

But really, we wanted to spend more time orchestrating many agents, orchestrating these powerful agents to abstract away the complexity of the files and of the changes.

So this is where we're putting our money.

This is where I think a lot of people are betting and excited about the direction of the industry, where the nature of your job, the nature of a software engineer, is going to change from standing inside of a traditional editor, typing lines of code, to supervising many agents and seeing hundreds of lines getting generated every second.

STEPHANIE WONG: Right.

Well, I'm excited for that future.

Well the future is here.

KEVIN HOU: The future is here.

You can already do this.

STEPHANIE WONG: So this is-- KEVIN HOU: It's only going to get better.

STEPHANIE WONG: Exactly.

OK.

So, Andy, what are we looking at here now?

ANDY ZHANG: Yeah, so in this case, we've created a heat map of all of the countries.

So in this case, the more blue it is, the more metals it has.

And so you can see and get a sense of the gradient of all of the different countries and seeing which ones have a lot more metals versus which ones do not have any.

Yeah, so that's one iteration of it.

There's many different ways that we can try to visualize this, such as another follow-up that I can imagine is like, OK, now apply this heat map onto a world map and visualize that.

So it's really up to the user at this point on where they want to take it.

And I think that's the magic of it, is that the colleague that is Antigravity will work with you and won't just go ahead and go rogue on many things, but instead it kind of constantly checks in and being like, hey, how is this looking so far?

Do you like this direction?

If so, let's just continue.

If not, let's course correct.

And that way it really feels like a partnership and you feel really in control of where you want to drive the work, whether it is the topic you want to learn or whether it's this new website that you want to build for the weekend.

STEPHANIE WONG: Yeah, I feel like now with this new way of building things, it's like, how do you know when you're done?

Or are you ever done?

I mean-- KEVIN HOU: Never done.

STEPHANIE WONG: --it feels like you're just constantly just you can iterate so quickly, so it's always a work in progress, but it's also exciting.

ANDY ZHANG: It is.

STEPHANIE WONG: I don't know if it's great for perfectionists, though.

ANDY ZHANG: Maybe not.

STEPHANIE WONG: No but I think it's a really exciting thing.

First of all, is there anything else you want to show with this because I have a big question?

KEVIN HOU: What do you think?

I'm excited to hear this big question.

STEPHANIE WONG: Well, I think-- I think-- ANDY ZHANG: You're getting us excited.

STEPHANIE WONG: What is the meaning of life.

No, I'm just kidding.

So I guess what's your takeaway for how we build things today using Antigravity?

It's not just software engineers that are using it.

And you talked about thinking, but I mean, it's not just coding tasks.

ANDY ZHANG: Right, it's not just coding tasks.

So much of it is like it really ties down to thinking, and a lot of it in my head is about being on the same wavelength as the other side.

And I think that if you meet someone, you usually take a bit of time to calibrating to figure out your wavelength.

But the more you work with them, the more they understand you, and like sometimes 30 minutes into a conversation, you realize you're just on the same wavelength and ideas just start coming.

And I think that's one of the most exciting things about this type of collaboration model is that it just takes a bit of time to identify that wavelength.

And once you really are on that same wavelength, and there's a lot of this context that you've talked about, then the ideas just start coming.

And I think that at that point in time, when you are on the same wavelength, that's when you really start surprising yourself.

You're just like, I didn't think of that idea, but this agent brought all these ideas out of me.

And you get impressed with yourself.

And it's almost an addictive feeling realizing that you're able to surprise yourself with a lot of these ideas, whether it's this app that you never thought you could build or I'm pretty sure there's many times when I was like, in school, I would be like, oh it'd be really cool if I could build this.

But you're trying to visualize the path to getting there.

There's all these difficult concepts, and you're like, I don't see the path there.

And then suddenly with Antigravity, I think, oh, I would love to build this.

But also walk me through how you would build this.

And I want to understand the different pieces, as opposed to just building this.

I want to be a part of the process.

I don't want to just see the outcome.

So it ends up being a really fun journey, both the process and also the outcome.

And it always brings me back to square one where I'm like, I want to restart this and try to build something new over the weekend or in the evenings.

STEPHANIE WONG: So if you are just trying to solve a problem, it doesn't even have to be a traditional software development problem, consider Antigravity.

Because if you're just using it for coding tests, you're under-utilizing it.

ANDY ZHANG: Totally, yeah.

I think that the takeaway for me is that obviously everybody only has 24 hours, unless you're some sort of super human being, per day.

And I think that in the time that I feel like previously I spent a lot of time on Google Chrome where I would browse and research things, search for things, but I found that increasingly I've been using Antigravity a lot more, and it makes me realize that I've shifted a lot of my time from Google Chrome into Antigravity,

mainly because I feel like within Antigravity, it has the tools to fetch a lot of information through Chrome.

And obviously, I still use Chrome.

Chrome is really, really great for doing specific types of research or testing out web apps, but I've really enjoyed using the Agent Manager with Antigravity to do research, get work done, iterate on things, review work, and also just collaborating with all my coworkers.

KEVIN HOU: An interesting rule of thumb that I try to keep in my head is if I do something more than, let's just say three times in a week, that is something that the agent should do.

And that way of thinking, not necessarily with code, sure there are going to be repetitive coding tasks, but I think people are already in the frame of mind to use agents for this.

But even I was talking to Paige a few weeks ago and we were talking about flight checking, and travel planning, and these sorts of things.

If you get in the mentality that an agent is very good at boiling the ocean, it's really good at going out, seeking information, bringing it down and condensing it into something you can understand, that mentality will take you very far for not just coding tasks, but really anything in your life that involves this sort of synthesis of information.

Another example is I'm planning a trip, and not only are there logistical considerations with this trip, but as a photographer, I want to see what are the places I'm going to.

What are some example photos that you can get at each of these locations?

That's not going to be very well summarized inside of a traditional Gemini app interface.

So what I do instead is, OK, Agent Manager, very good at making visual artifacts.

So go out, make me a website, pull in all those different images, make that travel itinerary, but make it an interactive website.

In this way of thinking where suddenly this previous task that you might have not considered to be part of the scope of what the agent is able to do, if you let your mind be a little bit more flexible, it can really do a lot outside of the scope of just coding.

STEPHANIE WONG: Yeah, I just visualized a future where the modes of how we interact with the online world are going to change.

KEVIN HOU: Totally.

STEPHANIE WONG: They're going to be very personalized, very customized, whether it be from yourself building it, or the tools that we use are just serving you extremely customized experiences-- KEVIN HOU: Totally.

STEPHANIE WONG: --like that.

So it's exciting.

I think you bring up also another good topic, which is how can you think about using Antigravity versus using, like you said, Chrome or Gemini app?

I think a lot of people might be like, well, I use Gemini app.

It's more familiar to me, how do you think about those two?

KEVIN HOU: Yeah, yeah, I think in the short term, what you should consider is Antigravity has access to your computer and your file system in a safe and sandboxed way.

But what this means is let's just say you're doing interactive data science type of activity or something that requires maybe a CSV that sits on top of your desktop or a file that you just downloaded.

This is the type of thing that I would probably lean on Antigravity to do a bit more.

Anything code adjacent I think makes sense to do inside of Antigravity.

But the Gemini app is great for a variety of other use cases.

You're on the go.

It's mobile.

There's a bunch of different strengths.

And I think the nice part about, frankly, what we're doing at Google which I find really, really exciting, is that these products are all unified in the sense that they're using the Gemini model.

When the model gets better, all these products will get better, and we're all contributing to that and making sure that you get the best possible experience across all the different Google products.

So in many ways, there are probably more overlaps than there are differences when it comes to using these products.

STEPHANIE WONG: Yeah.

ANDY ZHANG: And I think one thing I want to add on top of that this is related to the design philosophy around the Agent Manager.

So historically, there's also the editor feature.

But I think that when folks currently install Antigravity, that's the first place that they get plopped in.

But increasingly, I think the Agent Manager is the starting entry point, and you can see from the Agent Manager over here that there's actually very little that is technical, that is on the surface of the screen.

It is very much designed to be approachable by anybody.

You don't need to have a CS degree to understand what is going on in the screen.

And I think that that's tying back to the earlier concepts that we were talking about, which is like, what does an agent-first experience generally look like.

And also, how do you provide it with all the tools that Kevin was mentioning with having a sandbox file system?

So I think that having something that is extremely powerful but very approachable for everyone, but also has the capability to go that last mile.

Like in this case, on the top right, there's this open editor feature if I want to do that, those finishing touches.

So the way I think about it is agent manager can do like 80% to 90% of it.

And it also has the extensions needed so that you can wrap up and do that extra 5% to 10% if you hop out into the editor.

And with that in mind, that basically opens up the doors for the Agent Manager to everybody, because suddenly you don't need to have this knowledge or that prerequisite knowledge to use this.

You can just go ahead and start using this similar to how you use Gemini app.

The main upside, as Kevin mentioned, is that there's more toolings.

And there's also the agent, more capabilities and specific aspects given that it's a local environment.

And with that, you can start using it as the entry point for a lot of your work.

STEPHANIE WONG: Yeah.

And, there are a lot of tools at our disposal.

So I think it's just really exposing yourself, just trying it out and seeing what the limitations or possibilities are, and what is more useful to what you're trying to achieve.

I think, actually, we are getting some questions here on YouTube.

This person is actually wondering why is Google AI Studio different than Antigravity?

Are the models the same on the back end?

So I think there's some similar questions, like how can we think about AI Studio and Antigravity as well?

KEVIN HOU: Totally yeah, the AI Studio team, we work very closely together, actually, and a lot of the underlying capabilities of the agent, you can expect to be very similar, not only because the model is the same, but also because the way we've constructed the agent is this unified experience across the two products.

And what we'd really love to strive for and we're getting there closer and closer, we're shipping a bunch of great features here with Logan's team.

Making sure that AI Studio and Antigravity can play very well together, so what that means is, AI Studio is an amazing way for people to build really quickly a first iteration of what they've dreamed of.

There will reach a point, and maybe this is the topic of this podcast, in general.

They would reach a point when perhaps you need to get a little bit deeper into the code.

Maybe you want to bring it onto your local machine.

And that's really where the handoff between AI Studio and Antigravity will come into play.

There are different sets of features, sure, but I think the biggest thing to think about is you're going to be iterating in very quick succession in a web environment on a website.

If you want to do different types of projects, or if you want to actually get into the file system itself, this is when you'll bring it onto your computer and use Antigravity.

STEPHANIE WONG: OK, yeah.

I think that's a great way to think about it.

One of the things that we've touched on in the first section here is also just like the interaction modes you have available to you in Antigravity, which I think are a big differentiator for it.

So how can you think about, I mean, we've talked about how we're making it more intuitive for people.

Can you just tell us about how you think about that?

Why are we so focused on interaction modes with Antigravity?

KEVIN HOU: So for interaction modes, the most common mode that comes to mind today is the chatbox.

And I think that historically most folks, most people all know how to interact via chat.

You use iMessage to text your family or friends, or you might use Messenger or any other messaging app.

And so that was a pattern that people were already familiar with.

And so when we were thinking about building Antigravity, the thing is, how can we remove as much learning as possible so that you can just go ahead and it's like you already know how to use it?

So I think that was the first thing.

But when we want to think about going beyond that, the question comes to mind, how do you empower the agent to do even more?

And so at that point the question becomes is the chatbox enough?

And so similar to how we were talking about working with a colleague, one way you talk with a colleague is you just tell them, hey, how's it going.

But there are many other ways that you can communicate with them, whether it's through body language, through GIFs, through all these different ways of communicating.

We're not limited purely with words.

And so when we were thinking about interaction modes, many of the subsequent ideas is around how can you be more specific with the context that you want to provide?

How do you specify the area?

How do you specify what part of the image you want to comment on?

How do you give context by interacting with the editor?

So one of the really cool things that we do within the editor is that we understand which files you're generally looking at, what you are doing in your terminal, so that when you have a question inside of the agent, it already knows what you were trying to do in the terminal.

So as an example, if I was trying to run a Git command and I accidentally forgot, if I asked the agent, it knows that what I tried.

And it can infer, oh, this is what you're trying to do.

And so there's many of these different types of signals, whether it's implicit or explicit, that the agent can gather, so that it can provide more value to you.

ANDY ZHANG: I'll add one thing which is just ultimately, we need to build a product around agents.

Agents are going to do more.

The way that people interact with it is going to change.

And this is one of the most important things that we can do as a team, to enable people to use agents.

It's not as agents are going to do more work, for example, what are the touch points that you as a user will want.

If you're only interacting with something every 10 minutes versus every 30 seconds, the interaction pattern is going to be very different.

And so we focus a lot on interaction, because that is really the thing that is shifting underneath our feet.

These agents are able to do more, but we need to enable our users to actually have control and work well with these agents.

STEPHANIE WONG: Yeah, agreed.

And I think there's this shift.

As you mentioned, you have repetitive tasks throughout the week, and you're like, OK, now I'm shifting my mindset to saying, oh, let me think about how I can enable an agent to do that instead.

But that's already, I think, something that people have to get used to.

And so it seems like you want to make it as easy as possible for that transition to happen.

And so interaction modes is like, OK, how would a human interact and say, let me hit up my buddy.

If you had an assistant or something, you'd be like, oh, maybe they can take it on.

How do we get that mindshift to happen with software instead?

But it's interesting because I feel like, as you said, the way that we've started with AI applications is through text and chat.

So we've actually become so accustomed to that even in the span of one year, that now people have to almost relearn or refocus on actually, no, we can do it the way that humans would.

So I find it interesting.

It's like we started with human interaction.

We moved towards browser and mouse and key.

And then we go to chatbots.

And now we're going back to that intuitive human interaction mode.

So it's, I don't know.

It's an interesting shift.

ANDY ZHANG: It's so interesting as a challenge.

The other thing that came to mind is, if we zoom back to maybe 5, 10 years ago before agents were even a thing, most of our interactions were all at the editor level.

So it felt like as if depending on your level of abstraction, how you interact evolves.

So initially, the level of abstraction is just modifying code directly.

And you can only interact with something that you can understand.

And so I think that kind of alienated a lot of people who didn't understand.

And so as you zoom out of the abstraction, one level higher and higher, that opens up the doors for a lot of folks.

And so obviously, if we work with the constraints that the interactions are always approachable, and that as abstraction gets higher and higher, also conceptually is also approachable, that suddenly you have something that is usable and also approachable by everybody.

And I think that would be an amazing world to have where the Antigravity can be used by everybody to do whatever they want.

STEPHANIE WONG: Yeah.

One of the interesting conversations I've seen happening online from folks on Twitter and people at Google, it's like, we've seen these amazing capabilities with AI software development tools, but they've actually maybe helped more experienced software developers more than junior, because you can actually get that very last 10% where you're like, OK, I've got to debug.

I can actually go in and understand exactly what happened.

And I have that perspective.

I think that's changing over time, of course.

What's your take on it in terms of enabling just everyone with higher levels of abstraction, more capabilities, more advanced ability to debug to that last 10%?

KEVIN HOU: Yeah, it's tough because what the agent is good at is very different month over month, quarter over quarter.

I think it's actually interesting talking to people about junior engineers who are thinking about what their career should be and whatnot.

I actually do think these tools can make it such that everyone gets a huge lift.

And it's a question of who embraces it, almost.

If you are just starting out, suddenly your baseline for what is considered an effective engineer is so much higher, and that can be very daunting.

But at the same time, that also means if you embrace these tools, you can become that much more empowered.

And so there are different ways.

They might harness the fact that you can get from 0 to 90 as a junior engineer.

But a senior engineer, like you said, will take that last 5% and really know what to do with that last 5%.

So the agent has a way of meeting you where you are.

And I like to think that at all experience levels from someone who's never touched code to someone who's been soaking in the industry for 20 plus years, the important thing is that you buy in, that you embrace, and that you are willing to learn new tools.

And I think that even the way that we were coding a year ago, in many ways, because things are changing so quickly, it is almost an even playing field, where a year ago, the skills and tendencies of reaching for AI are so different than what they are now, and the person who is able to embrace that is going to be competing at a much higher level.

STEPHANIE WONG: Yeah, and I can guarantee you that anyone that's watching this, I think, is in that bucket of people who are embracing it wholeheartedly.

KEVIN HOU: Totally.

STEPHANIE WONG: So it's really exciting to see.

I do want to bring in a couple more questions.

We have a little bit of time left, so would love to see what folks are saying online.

Let's see.

OK, I'm going to double check what we have.

So we have a question here from LinkedIn from Adriana who asks, do you support versioning or rollback?

Can we go back if a generated change doesn't work as expected?

ANDY ZHANG: We provide a lot of optionality within Antigravity.

So obviously, we have a built-in IDE.

And so there's the version control tools that you love.

You can use Git and you can create commits.

You can revert that, and so all of that is still built in.

KEVIN HOU: Why don't you show them?

There's a button on each message.

You go into the conversation.

It's this thing called revert.

And basically, if you go up to your previous message say, I actually don't like what you just did.

Just click that button and you'll be able to revert up to that point.

So you don't have to click it.

We like what you did here, but you see the idea.

ANDY ZHANG: Yeah, so in this case, I can go ahead clicking revert per message.

And so basically, the way that the agent captures work is in between the messages like, OK, what are the changes I made in between since that message.

And then it creates a diff of all of those.

And then it allows you to revert it.

So that way, if you want to course correct from that point, because you felt like that was the point where you had a bit of a checkpoint.

You're like, OK, I want to maybe fork off of that and start adding something off of that.

I can go ahead and click over here, Undo changes up to this point, and upon clicking on it, it also shows a confirmation of hey, which files are going to be affected.

And then you can also view the diff for each of these.

And so if I click on Confirm, it'll go ahead and undo all of my work.

And so these are the different ways that you can course correct, whether it's moving forward and adding new changes or whether you want to move backwards, which is like reverting some of your changes.

STEPHANIE WONG: Yeah, yeah.

And that's really important, I think, just to give people like that peace of mind.

It's like, oh, actually, I don't like that.

Let's go back.

Yeah.

ANDY ZHANG: Exactly.

STEPHANIE WONG: Nice, OK.

So we have another question here from YouTube and they ask, how can you import projects from VS code to Antigravity?

Can you do that?

KEVIN HOU: Yes, you can.

It's a very similar model.

If you go back to that original screen, you can basically just, what is it?

Open workspace, and open that same folder that you would normally open in VS Code.

And it'll just appear.

STEPHANIE WONG: Nice, very easy.

Straightforward OK.

One more question here from YouTube, what other models can be added to Antigravity?

I know we talked a lot about Gemini Flash, which is awesome, but can you use other models?

KEVIN HOU: Yeah, we support models from OpenAI.

We support models from Anthropic.

And we support models from Google.

I think there's a lot of things that we could talk about in this world.

I think with Google specifically, we've built and optimized our product for Gemini, and you can expect the best possible product for the Gemini models to be Antigravity and vice versa.

But we do support a lot.

We want that open ecosystem style.

We want people to be able to try different models, so we don't restrict on that.

STEPHANIE WONG: Yeah, that's great.

I'll squeeze in one more here from LinkedIn as well.

So as we move into agent-based workflows, is Antigravity thinking about how to handle specifically permissions, access control, and boundaries for what each agent is allowed to do?

KEVIN HOU: Totally.

I'd say this is one of our P0's, if you want to use corporate speak, one of our highest priorities.

I think safety is super, super important for getting people to buy in and trust that these agents won't go off and do something potentially harmful to your codebase or to your machine.

And so we've been investing a lot of time to make sure that this is done correctly and done well.

What I will say is currently we have a number of guardrails and Antigravity is used by a number of teams. Enterprise, I mean it's used.

We have a different variant internally at Google, but a lot of the same principles apply.

And so obviously security is not something that you can just ignore.

And so we have things like domain allow lists, for example.

In the browser, you can specify which domains you actually want the agent to have access to.

You can specify which terminal commands you want it to run.

By default, it will ask you before executing any of these terminal commands.

We also have a sandbox mode to essentially make sure that it won't go off, and make harmful changes to files that are outside of your workspace.

Your workspace is defined as the project you're working in.

If you want to go off and access your Downloads folder, won't be able to do that.

So we've done a number of things in this way to just make sure that people feel safe with the app, and I can personally attest it does do that.

STEPHANIE WONG: Yeah, great.

I know we have a couple of minutes left, but I did want to touch on just like personal tips, tricks, and workflows that you use.

What are your favorite ways to use Antigravity in your personal workflows?

I know we've talked about, before this session, we've talked about planning mode or learning mode.

What are those things?

ANDY ZHANG: So a lot of my time these days, I haven't been able to get as deep in the weeds for code.

And I think I'm thinking a lot more about what are the next cool features we should do.

How can we increase the trust in the agent?

And I usually work very closely with the engineers, and sometimes it's important to be on the same wavelength, as I mentioned earlier.

And so what I like to do is I look at the code base and I frequently ask the agent like, hey, can you explain this.

OK.

What does this relate to?

And typically, if I did this 5, 10 years ago, I'd be going through every different file and trying to connect the dots.

And I think Antigravity is so good at connecting the dots for me such that I can understand things at the right abstraction level so that I can prepare for different conversations with whoever I'm working with.

And so it goes back to using it as a tool for thinking, and understanding which abstraction level I should be understanding, getting that knowledge from Antigravity, and then using that for the purposes of my work.

STEPHANIE WONG: Nice How about you?

KEVIN HOU: I'm probably spending-- I think my favorite way to use it is to spend most of my time in the implementation plan.

Especially with a code base that is as large as Google's, I mean, you could read about this.

It's probably the biggest code base to ever exist.

I spend most of my time, let's just call it 80% of my time, actually iterating on the plan.

And if the plan is pointing to the right files and high level has the right brushstrokes, I have much more confidence that the output will be correct.

I think if you did not have that, a much higher likelihood for going off the rails and maybe changing files, and then at that point, you're on a slippery slope where you might have to throw away all the work.

So I spent a lot of time iterating with the agent before it's even made any changes.

And I will do this for three or four different things that I think I'm going to do that day.

And then I will fire off those four tasks, either in parallel or in serial.

That part matters less.

More just I'm spending a lot of time planning, and once it has a good plan, I feel pretty confident that it'll be able to do the right thing.

STEPHANIE WONG: I feel like that's also a shift.

Of course, we want to see the magic right away.

We want it to see it build.

But are people spending enough time with the foundation and the thinking, and thinking about the problem.

What does success look like?

What is the plan if I were to think about it first?

What do you think?

ANDY ZHANG: Yeah, I think it's a really important stage, which is why our default mode is planning, because we want to make sure that this is at the forefront of what you should be thinking about.

And I think a really good colleague will not only do what you ask it to do, but it also tries to guide you through whatever is the best practice.

And so our agent usually will decide whether a task needs planning or not.

And if it does deserve a lot more thinking, at that point, it will ask you like, hey, let's collaborate on this and it'll show you warnings, and hey, this is actually an important thing to review.

Make sure you review this before proceeding.

And so that way you feel like not only can you trust it for building, but you can also trust it to tell when you should be thinking more versus less, especially if it's like a humongous refactor.

It's very hard as a challenge to even estimate how much work something can be, whether it's just a one line change or whether it's maybe not possible, or whether it requires a significant refactor.

And so with the understanding the Agent Manager has, it will help you figure out how much thinking you need to do.

So that way you understand the work that is involved and whether you want to move forward, or you want to change how you approach it.

STEPHANIE WONG: Yeah, and I think to wrap all of this together, going back to your original point of this is a tool that anyone can use.

Because anyone is a thinker.

Anyone has a problem that they can approach and solve and find success for.

So if you are a thinker and you're tackling any kind of issue, start with the problem.

Use Antigravity to iterate on that problem first, and then see what it can build and how it can do that.

ANDY ZHANG: Exactly.

STEPHANIE WONG: Amazing.

Well, we are just out of time.

So I just want to thank you both for the amazing conversation and for showing us the Agent Manager, Artifacts, and Interaction modes.

KEVIN HOU: Thanks for having us.

Always good to be here.

STEPHANIE WONG: Yes.

Well, thank you, everyone, for joining our live stream.

Once again, if you want to check out all of the Antigravity resources, check out the description.

Just a couple of quick reminders, we just launched GEAR, which is Gemini Enterprise Agent Ready.

So if you are interested in AI agents, they are the biggest platform shift in a decade, GEAR is your new home base to build and scale them.

From Hello World to production, you can get learning paths, earn official badges, connect with a community of builders.

Speaking of connecting, mark your calendars.

Google Cloud Next 2026 is happening April 22 through the 24th in Las Vegas, and registration is live right now.

We have an incredible lineup of people that you can meet and wherever you are in your developer journey, check out all the content, make sure to check out the registration link, and grab your seat before tickets run out.

And we can't wait to see you there and we'll see you next week on Google Cloud Live.

See you, everyone.

[MUSIC PLAYING]