Defining the agentic AI era

[MUSIC PLAYING] LOGAN KILPATRICK: Good morning, everyone. Welcome to day two of Google I/O. Hopefully, you all are as excited as we are about the announcements from yesterday, lots of cool stuff. This is going to be a ton of fun. I feel the theme, at least my takeaway, and I'm curious for you all as well about yesterday, was the agent era feels like it's really in full swing with Gemini. And so I've got lots of questions about the details of that story. And I'm joined by a bunch of wonderful, wonderful, wonderful guests. Josh leads the Gemini App team, AI Studio, and Google Labs. Koray is our CTO inside of DeepMind and Google's Chief AI Architect. Liz leads Search, which is an incredible responsibility and privilege. And Jeff is our Chief Scientist and also one of the co-leads of Gemini. So this is going to be a fun conversation. Maybe we can just start with Gemini 3.5, the agent era. Koray, what did it take to make that happen? KORAY KAVUKCUOGLU: OK. First of all, thank you very much. It's great being here with everyone, a wonderful environment 3.5 Flash, we started the 3.5 series with Flash. The biggest thing that we really wanted to focus on was getting coding and agentic workflows much, much better. We see a huge improvement from where we were six months ago, September, when we released 3.0, when the model had really high levels of reasoning and multi-modal understanding. And together with these agentic capabilities right now, I think we have a system that is really enabling this agentic era, as you said, for everyone. We are enabling that for developers. We are enabling that for businesses through Antigravity, through Gemini Spark. I think an exciting era is coming. A lot of focus went into really understanding those workflows internally, us using the model very, very intensely for coding and non-coding tasks as well. I think, to me, that has been the main driver of understanding where the real bottlenecks are and then getting that feedback into the researchers, into the model builders to extract that information and find the solutions that's going to make the models scale up, because there are hard technical challenges in terms of going long and solving long horizon tasks. But I'm very happy and proud for the team. LOGAN KILPATRICK: Yeah, no, I love it. The team did a great job. I think one of the threads that Sundar talks about a lot is this full-stack AI approach that Google has. And I feel like an interesting observation of that is that it feels like you actually need, in a lot of ways, this full-stack AI approach to pull off agents and actually make it work. And I think there's this symbiosis between-- we were talking earlier this morning about the symbiosis between the model, the harness, the product, and actually, Jeff, maybe the hardware as well. And you need hardware in order to scale this up. And I'm curious, any reactions to, obviously, the eighth-generation TPU available and powering a lot of these experiences in Antigravity, and maybe in the Gemini app, which is super exciting. JEFF DEAN: Yeah, I mean, I think we've now got a pretty significant track record of doing many, many generations of TPUs. And we've been refining the designs of those systems for-- initially, we were just focused on making them awesome for training and inference. And I think now we've got a bit of a separation of the designs so that we focus on, both training, which is a somewhat different problem, then inference which has reasonable splits in the actual chip designs for those two different use cases. And I think it shows in the inference speed of these models. That starts from having really good hardware that can run very large scale models very quickly, and with great interconnects and all these kinds of things. But it also is the focus of a lot of people on the inference serving team and the software stack above that to make this all come together. But we know when you use a model and it's super fast, it's a much more delightful experience than when you're using it. And it's slow. And that's true even if it's agents using it on your behalf. LOGAN KILPATRICK: For sure. Yeah, and Liz, I think maybe the Search team feels this more than anyone, the obsessive look at latency and getting users the answer as fast as possible. I'm curious, actually. I feel like there's a lot of balancing that Search has to do. You want to get users an answer quick. But obviously, Search has agents now, and it's becoming an agentic product experience as well. And I'm curious how you're thinking about finding that balance. And also, the headline of Search is, the biggest upgrade to the search box in 25 years. I love that. It's super exciting. LIZ REID: Yeah. I think it's been a really close partnership with Koray's team on this. One of the things that's been interesting, we've always known that latency matters a lot to people. But it actually requires quite a deep understanding of, how difficult does the user find the task? And how much value are you providing? And so people's willingness to wait is in part based on, how much work are you taking off? So I was talking with someone yesterday who was like, well, how much time can you spend? Well, it depends on what you're asking. If the question in the user's mind is quick, then you better be lightning fast. Otherwise, they're like, why the heck am I waiting? On the other hand, if the user would have spent 15, 20 minutes doing it, you can have 10 seconds, for sure, if you can do something amazing. And so we actually have been continuing to evolve over time what is the amount of time required. And then you take something like we talked about, bringing agentic coding in. If you're going to create a weekend planner for me that I'm going to use for the next several weeks, I might be quite happy to wait a minute even, because I can set it up. I can go do something else. I can come back, and I can use it. And so we really need to build that level of adaptation with something like information agents that's working behind the scenes. So you can think about latency in a totally different way. It's important that when the change happens, you're really quick to respond to the user. And so it's constantly going to this question about, what is valuable to the user? And as the models get better, that's great because it used to take a long task. And now you can do it with a smaller model. That's really exciting to see. And then at the same time, you come up with harder tasks to give the larger size model constantly. LOGAN KILPATRICK: Yeah. I feel like there's so much alpha in trying to find this right balance of, how long should you actually take to do a task? And Josh, I asked you right before we got on stage. Yesterday, that the Gemini app launched to trusted testers and hopefully for Ultra subscribers-- next week? JOSH WOODWARD: Yeah, next week. LOGAN KILPATRICK: Next week. Gemini Spark, which is the always on, 24/7 agent. I was asking you, what are the tasks that you have Spark sort of kicking off for you? I think we're all going to be trying to answer this question of, what would I be willing to let the agent go do on my behalf for some extended period of time? What are the things that I'm like, actually, I kind of need the answer right now? And I'm curious, for Spark, how you're thinking about this for the Gemini app, in general, but also any insight for Spark, specifically. JOSH WOODWARD: Yeah, it's fun. As we've been testing Spark, it's one of the things we've been finding. Each day you use it, you find more things it can possibly go do for you. And that's what's really exciting. When I first started using it-- this was way back-- I was doing regular scheduled things. So every morning, it might go through a bunch of stuff for me. Then I started doing a bunch of things where I would set up triggers. So for example, if I got an important email from Sundar, label it P0 and do a bunch of research about his question and prepare the draft. Then-- LOGAN KILPATRICK: Don't send the email! JOSH WOODWARD: Don't send the email. LIZ REID: Very important, don't send the email. JOSH WOODWARD: And it respects that, which is good. Then I've gone through a phase where there's a lot of information-related things. For example, I'm a huge Oklahoma City Thunder basketball fan. And I have it where I get just a daily-- this was a nice surprise, actually. I'm getting updates anytime the team has an update. But it's talking to me in a way like a die-hard OKC Thunder basketball fan would talk to me. And so it's almost like you're, I don't know, at the pub or something chatting about your basketball team. There's all kinds of other things we're doing on the team. And I know we've talked about this one, too, where it'll go out and create documents for you or slide decks. That was some of the stuff I showed on stage yesterday. That's been really helpful too. Usually what I'm finding is that the first draft is quite good. And then it gets me started, and I can go edit it and do revisions as needed. So it's quite fun to see. I can't wait for all of you to try it. And, yeah, we're going to try to scale it as quick as we can next week. So, stay tuned. LOGAN KILPATRICK: Yeah, I love it. I feel like it's one of the challenges of these new capabilities landing. It's actually deploying them across the product services. And Koray, I'm curious the tension you feel. We built a great model, and I think, historically, it was not easy. But it was a little bit easier because there was a few products that were using it. And I think now the success cases, we take our models. We build them. They're built for all of Google's products and the external ecosystem to build on top of. And there's thousands of people now who are involved in taking the model and delivering it across Google. There's also thousands of people saying, I wish the model would do this thing. And from a model perspective, Koray, I'm curious how you're finding the balance of, where are we going in the world where everyone wants Gemini to be everything, how to find that balance. KORAY KAVUKCUOGLU: I also have my own wish list of-- which seems like it's always the least priority one when I talk to the team. LOGAN KILPATRICK: You got to talk to the team. Talk to [INAUDIBLE]. I'm sure she's here. KORAY KAVUKCUOGLU: From my point of view, I think it is important that we actually get all this feedback. Yes, is it a challenge? Because with where we are, when we release a model, when you look at what is happening right now, we released Gemini 3.5 Flash, Gemini app, Spark, Search, and Antigravity and API. So this is reaching pretty much everyone in the world, on day one, immediately. So, yes, of course, you end up making some decisions in terms of making sure that there are minimums that you want to satisfy all the way across the surface. But it is important that, at the end of the day, this is always a journey for building intelligence. And the steps that we take there is what everyone is asking for in intelligence, whatever surface it is. So there is that common factor, that either it's Gemini app or Spark or Search. But every one of us want and what every user, developer, or businesses want is more intelligence. And that is common. And that's the main driver for everything that we do. That's always the highest priority. And that intelligence comes in different forms. Of course, reasoning capabilities is important. Understanding the world is important. Agentic actions is important. So our first priority is get the fundamentals right. And then, together with that, make sure that the model is customizable. The model is interacting with the user in the right ways. I think that kind of strategy where we look at the model as a product itself, I think that is really important, so that we work with all the internal and external products and interfaces to develop the model. But that's the reality. We are in that in between of not academic research anymore. It is not a real, real product. There is no innovation. There is not a stale product. It is in that really exciting space of, it is a technology that is going out in the world. And you need to find that balance where you both push the technology forward and make it useful at the same time. LOGAN KILPATRICK: Yeah. I feel one of the points of tension that people always say is-- and actually, it's something that makes these types of events like I/O really difficult to have empathy for the model teams-- to echo your point, I think building the model today is not traditional software development, building a product development. It really is research. And sometimes it's research. You should be taking bets on things that maybe aren't going to work. And we don't know how it's going to play out. And it makes delivering a model on a specific day to the entire world an exceptionally difficult problem. And I think we pulled it off, which was awesome to see. And actually, we pulled it off, Liz. Something that I was surprised and super excited to see was 3.5 Flash actually available to everyone inside of Search, which I think is-- I don't know if that's the first time that we've delivered our frontier model to everyone. But I'm curious also what the increase in intelligence is enabling for Search, that you actually-- it's changing the way that people think about-- I think the tagline which I love is, "Google Search is AI Search," which is awesome. LIZ REID: I think we're seeing with 3.5 that the reasoning and instruction following really helps. You think of Search, you can think of it as one thing. But behind it, there's often a lot of different tools, for instance. We're asking the sports back end. We're asking the finance data. We're asking local. We're asking travel. And so you want to pull all of that together into an experience. And so as the model gets better at reasoning, it understands how to think about those better, how to pull the information together and make trade-offs better. I think one of the things that's been really great about being able to make progress in Search is collaborating back and forth with DeepMind, where we would go and say, OK, well, here's the problem we're facing. But rather than, could you please make this specific thing work for Search, it becomes a brainstorming back and forth between the teams about, what is the underlying thing? What are we learning about what's challenging about tool use broadly? OK, how can we partner to go make that right? What are the right types of evals that push the limits? And so that we try and find a way that you expand the model in a general way for everyone, but understanding Search's needs or Gemini app's needs, where we're finding the opportunities or the challenges with the model and pushing it forward. KORAY KAVUKCUOGLU: I think that is really the key, that I think we ended up establishing that relationship both with App and Search, that we try to go into the roots, I think. Let's put it really well, that you can't fix individual problems. You always need to go to the root cause and try to solve those. LOGAN KILPATRICK: Yeah, I love that. One of the interesting threads, Jeff-- you and I were having this conversation probably three or four weeks ago-- as the model progress continues, where the new bottlenecks arise. And I think you had an interesting perspective on this that I think is something that folks should be thinking about as they build with the tools, as they use the tools. So I'm curious for your thoughts. JEFF DEAN: Yeah. So I mean, I think one way of viewing agent-based work is the agents are going to be built on powerful models that hopefully are running on super fast hardware, and themselves are lightweight enough to be fairly fast and capable. But if you start to think about maybe new techniques and new hardware that make the time in the model actually shrink dramatically, then all the tools that these models are trying to use, many of them are built for human level interactivity, like list the files in the directory. Go get this file and open it up and do stuff. And if you make the model infinitely fast, Amdahl's law says if you're spending half your time in tools, you're not going to get anything better than 2x speedup in that, even if you had magical hardware that doesn't exist. LOGAN KILPATRICK: Which I hope we get. I would love magical hardware. JEFF DEAN: So I think we're going to have to, as we start to use agents for more and more things and build out a broader suite of tools that these agents rely on, those tools themselves are going to need a lot of attention in terms of, how do we make those as fast and robust as we can? How do we have maybe approximations to what the tool would really want to do? But maybe it can do a lighter weight form of whatever it's trying to do that is 10 times faster because I think that's going to be really important for building. If you start to have long-running tasks that write this code and it's going to take a week, you'd really rather have that code take half a day to do. KORAY KAVUKCUOGLU: There's a really nice cycle here. I mean, we are in a really nice position of inside Google, it's so automated, the whole engineering environment, thanks to Jeff and others, of course, creating this environment. But many things are designed and built for human frequency and human cadence. So it's an initial amazing step to be able to have this. But now we are getting into this exciting stage that we can actually now use our own models to actually go through the stack and adapt them to this new agentic world. And then tools get faster, and then the models work better. And then that is another cycle that I'm excited about. JEFF DEAN: Yeah. I can give you a good example, actually. We had a bunch of internal tools that the models were trying to use. And many of them were written in Python because that was easy and a good way to do it. But Python is not the speediest of programming languages. And so even the startup time of the tools was actually quite slow. But the models are now capable enough. When you're asking a model to write code with a natural language prompt, that's ill-specified in many ways because the model has to fill in a lot of details. But if you ask the model, here's a Python program that does something with a whole bunch of tests that test its functionality, please translate that entire system into go, that's a much more fully-specified problem because now you have, actually, a fully-functioning spec in Python. And you just want it to be in another language. And people have been able to rewrite many of our internal tools to be 10, 20 times faster with just one night of work with a capable model. JOSH WOODWARD: And I think one other interesting thing about this is there's the part of, how do the tools get faster and making the tools run at machine speed, maybe not human speed. The other side of it we, found building new products in Google Labs is around the context. How do you prepare the context in a way that's not just human-readable, but machine-readable? We have entire teams now that haven't written PRDs in months because that's the old way of doing it. You're writing, in a way, whether it's an MD file or other things, where basically one of the models can just pick it up and go with it. And even, we showed yesterday on stage, Stitch is another example, one of our lab's experiments. They've kind of introduced this whole idea of a DESIGN.md file, which is now open source, where you can just codify, here's the design language for our app. And the model can just take it and run with it and obey it. And so I think it's both the tool velocity, as well as the context. Making both of those human and machine-readable is really interesting. LOGAN KILPATRICK: Yeah. Were you going to add something, Koray? KORAY KAVUKCUOGLU: I mean, I think all this transformation, I think, is going to really transform. The way we do software engineering is changing, and we see this inside Google. I see this as an opportunity for us to also understand what are the next frontier of developments we need to do in the model. This is why I'm always excited about working with products, external, and internal as well. That first-party experience that model builders and researchers need to have, I think, is irreplaceable. That is always the main driver. So I expect a lot of folks here as well are going through similar kinds of things. Like, why coding is important? I mean, we test the models that way. But in a way, a big part of the transformation that we see today that is going to happen is transforming our infrastructure that is very much designed with the constraints that we had until now. And with this new agentic world, we can actually transform all of our software infrastructure in a much faster way to adapt to the new world. And I think that is going to be one of the biggest drivers of agentic coding. LOGAN KILPATRICK: Yeah. Speaking of agentic coding, Koray, you mentioned yesterday in a conversation that we were having, that, actually, you personally reviewed all of the DeepMind code before Google-- KORAY KAVUKCUOGLU: In the early days. LOGAN KILPATRICK: In the early days. You were personally the one reviewing every single pull request and change-- KORAY KAVUKCUOGLU: Yes, and at some point, I realized that doesn't make sense. But I did that for a while, yes. LOGAN KILPATRICK: Yeah. I'm curious, through that lens and obviously thinking, your job is very different. And actually, all of your jobs have changed in different ways over time. I'm curious for you all personally, how this advent of AI tooling and the way that we all work has shifted what you do, how you spend your time, if it has at all, or if you're still sort of operating in the same way. And I don't know who-- if anyone wants to go first or feels strongly about this. But it feels like the Renaissance is happening. And for myself, I was doing a lot less coding before I joined Google. I was doing a lot of coding before I joined Google. I stopped doing coding for a little bit. And then actually, this whole AI coding revolution has brought me back to writing software every day, which has been really cool. KORAY KAVUKCUOGLU: I feel the same. I used to write a lot of code. And then my career took me in a path where it slowly degraded and not much. And then, nowadays, I feel like when I see something, I'm like, OK, this could be done in a different way. Or I have an idea. I want to try. It is easy for me to do it because, just like I'm having all sorts of conversations on different things that we are doing day to day, I just have conversations with agents about, OK, I'm trying to change this part. Can you start looking at it? And I don't need to-- I can do it very asynchronously. That is what is enabling me, because even if my time is split between many things, at night, I can come back to it and say, OK, what did the agents do and see I like the result. And I love that. I mean, that gives me a more productive feeling, to be honest with you. LOGAN KILPATRICK: Who's reviewing the code, though? Is it Jeff? No, that's the real question. Liz, sorry. LIZ REID: I think one of the things that's interesting is just to watch the blurring across jobs because there's a lot of cases where people have some familiarity. But the startup cost is so high. OK, I want to understand this data question. I don't know where the tool is located previously. So then I would go bug the data scientist, who should really be doing a more advanced task, to please find this information for me. But why do I need to do that? It's just because I don't actually know where the data is located. What's actually the case? OK. Well, now I can just go work with the agent. And it can do that, and now I can do this. OK. Can I lean on the product? Can I change code? I hadn't written code for several years. Now I can write code again. Because a lot of the startup costs or a lot of the knowledge about where something was, that was difficult to be good at everything, especially in a code base as big as Google or Search. But you had the ideas. And I think you see this at my level, but you see it at the levels below. Instead of a PM trying to then go talk with the UX designer that doesn't quite look right, and then trying to figure out how they explain this to the engineer, who then tries to code it, can the PM change the skill or the design file, see it, realize their idea was not what they wanted, try again? And when they finally get something good, then they can go work with the engineer to make sure it's performant enough or other things like that. And so I think it's amazing because it empowers people to take the parts of their job that might have been a small percent but that they had to rely on somebody else to do and unlock it so that more people can feel like they're builders again and not just coders or not just designers or not ones. But they actually get to partner with other people to build something cool. JEFF DEAN: I mean, one thing I would say is it really accelerates the ability to do software development and experimentation, and particularly in parts of the codebase where you're not necessarily as familiar because you didn't write large parts of it. So sometimes I'll dive into some performance issue and start to look at, why is this particular thing taking a lot of CPU cycles across the fleet or whatever? And sometimes it's in parts of the codebase that I'm not that familiar with. But you can actually look at what the performance problem is and express at a very high level, oh, it'd be different if we used this particular kind of locking scheme instead of this one, or this particular kind of data structure instead of this other one. That would probably improve it a lot. But it's actually quite a lot of work for you to manually go do that change. But now you can just rattle off these ideas, ask the agent to measure the performance impact, run all the tests. And you can be much, much more productive. And as Koray said, you can do it asynchronously as well, which is quite nice. JOSH WOODWARD: Yeah, I'm finding that's one of the big changes. I do a lot more asynchronously now. And I really try to guard and protect the time when I do something synchronously. That time has gotten more valuable, I would say, just in my calendar. I also do a lot more on my phone. I also do a lot more with my voice. And so those are other areas, I would say, just things that you never-- maybe a project didn't feel possible before. but now you can even just do that project just on your phone. So I think those are things that are changes in my day to day. KORAY KAVUKCUOGLU: Maybe just a little plug. Spark is really designed for this kind of asynchronous interaction, as opposed to when you go to the Gemini app chat or Search, you expect to have this conversation ongoing. But Spark is really for these kinds of asynchronous interactions where the agent is going to go and do something for you. And then you can come back and check, or it will notify you that I'm done or things like that. I think that is a big enabler, really. LOGAN KILPATRICK: Yeah, this is a perfect tee-up for my question about the future of interfaces. And as the models change, as the capabilities change, we launched Gemini Omni yesterday and the ability to take in any input and produce any output, which is really exciting. We launched a bunch of new audio interfaces across the suite of Google products. Josh, I think the Labs team is often at the helm of trying to figure out, what do we do with these new capabilities? What does the next interface look like that's voice-native and solves all these problems? So I'm curious. I'm curious if you all are cooking anything that you can share, or just general comments. JOSH WOODWARD: Well, we are cooking a lot. We'll see if it's any good in a little bit. I mean, that's part of what we're trying to do, though, is figure out, if you take something, say, like, voice-only interface, and you dial that all the way to 11, what do you get on the other side of that? What kind of product? What kind of experience? Because at this point, I feel like the models with Voice-- we tried to show some of that yesterday, whether it was what Sundar showed with Docs Live, where you're just creating the document, or some of the stuff I tried to show with Spark, or even some of the stuff you'll see throughout the festival here on site. I think that's one area where we will try to go explore. I do feel like, just building maybe a little bit on what Koray said, so much of how we've used computers for the history of time has been synchronous in a lot of ways. And maybe it didn't always-- like, you type in. You get an answer in milliseconds. You may have to go make a coffee or something for some of the large systems and other things when they first started. But I think this idea that you'll be able to just toss things over your shoulder changes how we're going to have to think about the design. So our first design of Spark is a little bit more of a dashboard view. I don't know if that'll be the ultimate design. I'd be surprised if it was the ultimate design, to be honest with you. So I think there is going to be some notion. People talk about it. Is it a dashboard or a command center? Where are you orchestrating these things? LOGAN KILPATRICK: Mission control. JOSH WOODWARD: Is it mission control? LOGAN KILPATRICK: It's going to look like the NASA-- JOSH WOODWARD: Yes, that seems hard to read, though. LOGAN KILPATRICK: You've got 1,000 monitors and 20 people. JOSH WOODWARD: That's right. I do think, though, there is maybe an arc towards interfaces that do feel more and more natural, though. And I think that's probably also why I'm very excited about Voice, is because we know that just around the world, we've all grown up learning how to talk to people. And so I feel like that-- maybe there's some unlock here where Voice plus some notion of a status that gives you a sense of where things are or where they need your help. It feels like there could be something there. But we're going to try to experiment a lot. We'll see where it goes. JEFF DEAN: Yeah, I feel like the real the dashboard view is kind of a good one that will scale to a limited level. But it might be good for 10 things. But it's probably not going to be good for 100 things you're trying to do. So then do you introduce some sort of hierarchy of, these tasks are all related. Or maybe you have stand-ups among your agents or something every day. Who knows? I think that question is, if every person has 30 virtual interns working with them, how do you coordinate that? KORAY KAVUKCUOGLU: Or maybe we'll just do what Liz does, Search, and let the agent design the interface. LOGAN KILPATRICK: Yeah, that's right. LIZ REID: I mean, I do think there's this interesting question about whether the interface is the same for everybody. And actually, is the way that you think about processing the task different than the way I might think about processing or Jeff might? And so, if you truly have an agent, can the agent organize the people the way you want, not the way somebody else wants? So how does that affect it? JEFF DEAN: It definitely should be different, I think. I mean, in education settings, people learn in different ways. And so, a podcast describing something versus an interactive tool that people can play with versus a textbook chapter are all different ways of learning things. And they're differently appropriate for different people. LOGAN KILPATRICK: Yeah. One of the things that I would love y'all's reaction to, just because we're all so close to the action. And so I think it's interesting to hear your perspectives about anything that's surprised you that we launched at I/O or in the last couple of weeks or months. And I'll give my really quick one, which is something about Omni. And the part that surprised me-- actually, we were talking this morning-- I academically got it. | was like, oh, the model's powerful. It's Nano Banana for video. And then we had that podcast intro clip of what we were doing. And the example was a bunch of people talking in a podcast. And then I'm holding a cat and a plant. And there's things flying all over. And I was just like, it made sense. And it clicked for me. I was like, I get it. What I was surprised by is, that use case was what clicked for me. And that maybe is the application of taking videos and doing interesting things to them without actually changing the core content maybe to make it more engaging, Jeff, to your point, for educational context or add some supplemental stuff. So I'm curious for you all. Anyone have any interesting things that they've been surprised by from a launch model capability perspective? JOSH WOODWARD: Well, I've got one maybe real quick. Google Flow was on stage yesterday. It's a tool we've done for filmmakers. And one of the things that was kind of snuck in there-- and everything flies by so fast in the I/O stage. But we've got a new area in the product where you can make tools. And to me it could be00 we'll see if it works. But one of the really more big jumps in how we think about product design-- because we're essentially carving out an entire part of the app where we're saying, you as a filmmaker, as a creative using it, vibe code your own tool. And so it's taking advantage of all the 3.5 Flash advances and reasoning, coding. And people, I think, are going to be able to make actually to this point, their own software right in the tool. And I think for us, it was a big shift the first time I saw it as we were headed into I/O, because now as a product team, we're kind of building almost a thin shell with the space to let people that use the tool really make the tool their own. And I think that's very empowering, but also a whole different way to think about, I have to design every tool and every feature. It's actually, no, just put the tools on the table. And these are very creative people. They're going to go create the shaders and the text overlays and the lighting effects and everything they're going to want. And so I'm very excited to see how that hopefully takes off today as people started using it. Our Discord is kind of throbbing with cool, creative ideas. So that was one for me. LOGAN KILPATRICK: I love that. KORAY KAVUKCUOGLU: For me, the operating system demo that [INAUDIBLE] showed that we did in Antigravity. It was surprising to me. We worked on this model for a while. We work on these agentic capabilities, code capabilities for a while. But the capability of just saying, teamwork, go build an operating system. And then it just goes and hundreds of agents work for a day and a half. And they come back with something that is actually functional and working. Still, when I see that, I'm impressed. I'm surprised. I think I always, in a way, find myself in a position of, take a step back. To be able to do this is impressive. To be able to do this at scale is impressive. To be able to do this at this speed is impressive. And to be able to actually give this to everyone who's using Antigravity and being able to do it at that scale, I'm excited. LOGAN KILPATRICK: I love it. KORAY KAVUKCUOGLU: Yeah. LIZ REID: I think for me, we started looking at, how do we bring some of the agentic coding capabilities into Search? But initially, we were working with the model. And then partway through, we said, oh, well, maybe for some of these more advanced use cases, let's look at using Antigravity. And then that worked so well. They're like, well, actually, maybe we should use it for more things. Maybe we should use it for more. And I don't think, if we had told people four months ago, we're going to take Antigravity. And we're going to put it in the heart of Search. They would have said, what are you guys talking about? LOGAN KILPATRICK: Liz is going crazy! LIZ REID: Search and developer? These things do not go. But you do it. And you see how it unlocks the power to create these experiences that are very information at the heart, but transform them in a very personal and customizable way. And so I think we've only realized more and more how we can bring it in. And that's been pretty exciting to see. KORAY KAVUKCUOGLU: Just to open a little parenthesis there, when we say Antigravity, we mean multiple things. So Antigravity, there's the user interface that we give to the developers. But there's the harness. There's the SDK. We gave that SDK as well. So the same capability that Search is using, building on top of that SDK to build those agents into Search, we are enabling all the developers and businesses to use that exact same SDK that is really co-developed with Gemini to build applications for anyone who wants to use them. LOGAN KILPATRICK: Love it, yeah. JEFF DEAN: I mean, maybe I can tie together a few of these themes, because I think, collectively, they are a really exciting sort of vision. So in the past, software has been something that is created that has to be kind of standardized and then used in lots of settings, where maybe it's not the perfect fit for this particular setting or the perfect fit for that setting. But because the creation cost was very high, you settled on, this kind of does the job, and it's good. But I think now you're seeing, with long-running agents that can go off and create bespoke software, you can actually get very customizable systems that do much more closer to what you actually want in any sort of environment, be it your own personal setting, be it in the middle of Search. You get some nice visualizations that are created by the system, be it in tool creators for media generation. I think that's a really cool aspect of what these long-running tasks will enable, is you can just say, I would like this thing to exist. It will go off and hopefully autonomously make that thing come into existence. And then you can use it. LOGAN KILPATRICK: I love that. We have two minutes. So we're going to do a lightning round of fast questions. Jeff-- and I think you're uniquely qualified to answer this. The most important developer skill in 2026? JEFF DEAN: Learn how to use coding tools and agents to be much more productive and build awesome things. LOGAN KILPATRICK: I love that. Liz, something you hope that Search can take off people's plate in the next few months, year, that it doesn't do today? LIZ REID: I think for each person, it's probably a little bit different. But it's, what is the part of research and searching and gathering information that feels drudgery to you? And you feel like you can just get rid of that, give it off. And then you can spend your time exploring and learning about the things you want to learn about. LOGAN KILPATRICK: I love that. Josh, something surprising that Gemini Spark has done for you in the last month or so? JOSH WOODWARD: Ooh. It's very good at looking at your calendar and suggesting meetings you can cancel. And it will even-- [APPLAUSE] What's amazing about it too is it'll tell you why. And it can even suggest a reply, if you want to say-- LOGAN KILPATRICK: One-on-ones with Logan? Who cares? Cancel. Cancel, not important. I love that. And, Koray, maybe when do you think we'll get a Gemini Ultra model? No, I'm just kidding. [LAUGHTER] Obviously, huge amount of progress. What do you think comes next? KORAY KAVUKCUOGLU: What do I think comes next? I think, from the developer, user's perspective, a lot more agents, because I think now we are at the place where we are figuring out, what is the right way that we can actually surface agents for people to use in many different places in their lives? This is all about giving users the control, making sure that we get it right. I think we are going to see a lot more of this happening. And I think, from the technology perspective, again, I think we are going to see a lot more technology frontiers being explored, not just solving today's problems, but also pushing the technologies so that we are building intelligence. We are building something that is learning and thinking and understanding how to be creative, how to push the boundary. So I think those two things will happen at the same time. LOGAN KILPATRICK: I'm excited. I/O 2027, we'll see. This was amazing. Thank you all for joining us this morning. Thanks for those who are tuning in on the live stream. We're going to take a short break. And then James Manyika is actually going to be on the stage with our Google Quantum AI founder, Hartmut Neven. Hopefully, I said that right-- Hartmut. Join me in giving Josh, Koray, Liz, Jeff, a round of applause. KORAY KAVUKCUOGLU: Thank you. [MUSIC PLAYING]