Is Gemini Actually a Coding Model? Google DeepMind's Ivan Leo Answers

Abhi: If you were to do a prediction for the rest of the year, what's on your bingo card?

Ivan: I think we're gonna be rolling so many more agents in day-to-day life. I think we're just gonna see token consumption go up. My mom's asking me about OpenClaw now. It's really exciting to see a lot of people just start playing with models and feel the AGI, I guess

Shane: We have a guest, Ivan from DeepMind. Welcome to the show, dude.

Ivan: Hey, how's it going?

Shane: Thanks for being here.

Ivan: No, I mean thanks for having me. I think I've been a big fan of Mastra and what you guys are doing I think it's become incredibly exciting, especially in the space of building agents. And I think TypeScript has just been a language that's just perfect for it. And so yeah, really excited to be here.

Abhi: Now I know you came with a schedule of things that you wanted to share today. And before we get into that, because that thing is dope, by the way, the list that we're gonna get through. Can you give the audience a little background about who you are? How did you get to DeepMind, where you're from, which is interesting?

Ivan: I started out with open source actually. And so that was kind of how I started playing around with language models With a Python actually a Python package called Instructor where I think back then the problem wasn't that it was difficult to build agents. It was that it was difficult to work with language models reliably. And so that was my first foray into language models where I wanted to make it easy for people to basically be able to build applications on language models. And so I was a co-contributor to a package called Instructor. Fast forward like a year after that, I joined a startup called Manus, where we were building knowledge agents that were basically automating a lot of this knowledge work. Let's say you're a venture capitalist, a VC investor, you're a PE fund, you're a consultant. A lot of your job is trying to make sense of data, trying to understand like, I get this question, I'm trying to invest in this company. And so a lot of your job is trying to make sense of the information that's out there, pull it from related sources, put it together nice PowerPoints, or just, you know, at the end create an interactive visualization. And so I was there from, I think When they first moved to Singapore. And I was there all the way until they finally got acquired by Meta. And it was really exciting. I think I joined at a really great time. The team was great. Every night we were pushing great features and we were pushing state-of-the-art When I was kind of deciding whether or not to go to Meta, I was fortunate enough to get an offer from DeepMind to join the developer experience team. I ended up deciding to take that offer. And it's been really exciting since. I think My job is just to try to figure out how to make it easier for people to build on Gemini. It's a great model. And so that kind of is the background of how I ended up all the way from Singapore. To San Francisco.

Abhi: Welcome to the community. It's a nice place though.

Ivan: I feel I I've been very surprised by how friendly people are in San Francisco and how open and I think people are to just discussing very Technical concepts and everyone's working on something very exciting.

Abhi: They're more friendly to you if you have credits to give too.

Ivan: We're working on that. You know I think Maybe we can do a Mastra and DeepMind Hackathon.

Abhi: I think that'll for sure.

Ivan: Yeah, we should do that for sure.

Abhi: So big event happened what two weeks ago. I/O. All the bells and whistles were out. Could you give us a little recap of that?

Ivan: I/O is, if you're not familiar, it's Google's annual flagship conference where you know we announce all of the different product launches. For me, I think this was the first time I actually saw I/O and being there in person seeing the huge sort of venue and scale that it is was just really mind-boggling. So you have everyone from Waymo to the guys doing Android to the guys who are doing the Genie models and us who are working on the API and the models at DeepMind. It's just pretty incredible to see how far everyone's kind of pushing the boundaries. For me this year it was the first time I worked on a product and we launched it. It was called Gemini Managed Agents. And so What it makes it possible for you to do is that you can basically use the now the Antigravity harness in your own applications. We give you a sandbox out of the box and you can basically get it to Do whatever you need to do. But it was really exciting to kind of see a product that I work from start to finish, launch it. And then we had a new model And then just the general love, I guess, for everyone, for all across all of our different products. I think the overall token volume is something like 3.2 quadrillion tokens this month, which is I I never expected to hear a number that big. I think there were two I think two things that were really interesting to me that stood out initially at the initial keynote. One is that of course outside you're seeing this huge consolidation whereby In every single Google product, now Gemini is being used. I don't know if you guys have tried Ask Maps, where you can actually talk to Gemini within maps. It's it's my favorite feature. When I'm worrying about like, hey, today I'm I'm really hungry, what do I eat around me? I can just ask Gemini. Or hey I want to grab coffee with someone, where do I go in between us? It's just been really cool to see Gemini being embedded in a lot more of these applications. But also I think the really exciting part is joining DeepMind. It's very different when you see it from the outside and the inside. Because when you go in you see I know that was this post that went viral where Steve Yegge was like, Hey DeepMind's cooked or something like that. And then it was like no one's using AI coding or stuff. But it's actually incredible because internally I think the amount of tokens we're consuming is doubling every few weeks. I think it was something like a f I think we shared some internal numbers, but I think from the time that we announced that number, it's kinda doubled in like three weeks since then and it's only gone up. We're really trying to push on all fronts and it's it's just been really exciting to see And work alongside the people training the models, building the applications, and also now trying to make sure it's easy for people, builders who are using frameworks like Mastra, for example, to just be out there. Yeah. And use it.

Abhi: It's so cool seeing Gemini and all the surface areas that you use every day I was using Gemini and search, which then linked to Maps. Then I was in Maps and I was asking Gemini in there as well. Our team especially. We stopped doing Google search and we would just go straight to ChatGPT, right, for search. It'd like But now you can just go back to Google search and you can get the same kind of experience right at the top, right? And it's it's so cool to see Gemini everywhere now.

Shane: I definitely feel like Google search is is more back into my daily routine where it kinda dropped off. At least some of it had dropped off a bit, but it feels Back mostly to me.

Ivan: I think so. It's it's actually quite interesting to me. I think if you asked me two years ago, what was what was gonna happen to search engines like Google. I would have been like, oh I'm just gonna talk to Gemini or Chat GPT and I'm never gonna use the search engine again. But it's almost like the more you ask questions, the more Things you want to explore and then you end up really diving really deep. And I think AI Mode has actually been something I enjoyed a lot. There were some really cool demos I saw at I/O Where you know they chatted with on Google search where it was like, hey, I'm interested to find about black holes. And then actually created a whole canvas out of the box and it's all you could interact with it. And I'm quite excited to hopefully be on the way like the to be able to use that.

Abhi: So is Antigravity the new like that's where everything's going through, not Gemini CLI, right?

Ivan: Yeah, so I think right now what we're seeing is there's a lot of consolidation. So we have the Antigravity CLI, which I think is gonna be maintained a lot more. There's the Antigravity application, which is, you know, the sort of I guess the single application you can interact with. You don't need to see the IDE. I think it's a lot more people's Are kind of being comfortable working with models without having to expect every line of code. You know, that becomes a go-to. And then you have the Antigravity IDE, which is the original VS Code fork that they started with and then they forked it out. It is a little confusing. I think. You know, there's so many things that are happening all the time. But that's been really exciting. I mean, I think for me it's it's been something that I've been using a lot. But for me, I think one of my favorite new launches has been Omni. Do you guys follow this Twitter account called fofr.

Abhi: No, we should!

Ivan: fofr is someone on our team — fofr, @fofrAI. He's a prompt engineer in our team. He works a lot on the generative media side, and you can see he uploads A lot of different basic videos of what Omni is doing. One of my favorite demos, let me scroll down to find it. One really cool one is just you can see over here Google I/O. He like Replaced it with the Xbox. I kinda like screen. First Google I/O of 654321. And then there were another bunch of launches where You know, he took his feet and there's a guy walking down from it. All right. And I think the one that I thought was really cute was this one. So we have these I don't know if you see have these in your offices where You have a little Bevi.

Shane: Yeah, yeah.

Ivan: It's the one that gives you carbonated water and you kind of choose what you want, but he made it into one that spits out rabbits. And then I thought the cool one, let me try to scroll it down a little bit more. Was basically we have this podcast called Release Notes. So this is Release Notes. It's a really serious podcast where now maybe we showcase people talking, you know We talk about the new stuff that we've we've launched and you can see how Omni is able to splice in splice in these edits in a very natural way. I would love to have shown the audio is not working for some reason. You can see Logan with You know, the giant pot of plan and then you can see like, you know, everyone like the spaghetti.

Shane: Yeah, the spaghetti test is always, you know, I just remember a couple years ago, it couldn't You it was such a mess to try to get anybody any video to show eating, ex especially spaghetti, and now it handles it so so much better.

Ivan: And now Michael Jordan can be eating spaghetti for real. I think it's time to find a new benchmark, you know? And I think the funniest part was actually this was a Omni video that we launched. Initially we did a launch video where Logan said like Gemini and this was actually created by Omni. And so the really cool part of Omni is that it's kind of like Nano Banana for video. You can describe the edits that you want in natural language and it'll kind of like cut and splice around it, which is really cool. And that and the new Gemini 3.5 Flash model I have a few demos that I prepared ahead of time if you want to see what I created with Gemini 3.5 Flash.

Shane: Hell yeah. Yeah, of course. The aud the audience loves to see demos.

Ivan: So you can see over here, this is a Windows XP kind of clone that I got what is essentially 3.5 to do. It one shot this in Antigravity. So these are websites that I've created ahead of time that are hosted on Convex. And so you can see this clippy over here talking to you, it's moving. All these frames are created by Nano Banana. There's a Tamagotchi over here. Let me zoom in on it. And you can see basically all these frames are kind of being cycled through. And then this is for me just a simple habit app. But it's been really fun to vibe code and build a lot of these I think 3.5 is just a really creative model and it's been fun to just be able to do this. I've heard there's a test where you kind of are able to close this window all the way here and it should work, but I think I need to tune this a bit more, but Yeah, just been vibe coding a lot of these simpler apps and they've been really fun to do.

Abhi: That's awesome. Which Antigravity product were you using?

Ivan: I was using Antigravity IDE because I still like to look at Code. Yeah. But I I think I've been slowly migrating over to the new Antigravity. The one that you don't look at the code, just a normal Antigravity app Yeah, I don't know. I think for me AGI was when I used Gemini 3.5 Flash to buy everything for my new room. I just give it a browser tool and it's just go on Amazon, buy it for me and it bought a new monitor stand, it bought a bunch of lights, bought a smart plug and I'm I love it.

Abhi: That's legit.

Ivan: Yeah.

Abhi: Do you have a demo of the managed agents?

Ivan: Yeah. I think I did a bunch of demos over here. So what I did is I Got it to basically create a podcast about Mastra. So the way it works is that if you go, if we load it from scratch, you can basically tell it, give it a topic or creative outline. So for example, you could say like write me a short script without The Mastra team and what the framework that they created does And if you generate this, it's gonna boot up a agent in its own sandbox and this agent has runs the full Antigravity harness. You can add in additional skills, we'll have support for custom tool calls at some point. And right now this you can see the sandbox is pretty snappy and it's actually persistent by default. And so whatever you install on it, whatever you save on the f on the sandbox, you can kind of go back to it at some point. If you switch back over here, you get a very nice...

Shane: And unfortunately, I don't think we can hear the audio. Oh no.

Ivan: Yeah, yeah. I think I can I can be sure. But you can see this is kind of just using our own native TTS model. And so it's going back and forth. It's talking about Mastra. I did some web searches about you. And then from there, it's able to put together this NotebookLM clone. I never tried NotebookLM until like a week ago. And I've been using it on the go, and so I kind of want to replicate it for my own stuff. Because I don't want to log into the app and get it there. So I was just like, oh, let me create something with this agent.

Shane: Yeah, I was gonna say it looks, it reminds me of NotebookLM and you just, you know. Quickly build it, which is awesome.

Ivan: Yeah, so it's it you can do it all with a very simple API call, but I think definitely when you want a bit more control, then you gotta fall back to a framework like Mastra. Then you want these other models. But I think it's just a pretty exciting thing just to be able to take a whole product from the initial, like, hey, I'm trying this. This is a new product launching to take it all the way to I/O and have it there. So yeah, it's just been pretty exciting over there. Yeah. Have you guys been able to try the new Gemini 3.5 Flash model, et cetera?

Abhi: Well we signed an NDA, so maybe I'm not allowed to talk about it. I think mostly for us, we want to use m like for internal use. We're trying to do things that max on SWE-bench, right? And I feel like Gemini is never talked about as a coding model, or at least that good of a coding model. What would you say to that, I guess? You know, being from DeepMind?

Ivan: With regards to Gemini not being a coding model, I think for me I've often been very surprised at the kind of creative solutions that Gemini gives me. So for example, I was working with some Python scripts and when I used Opus and I used GPT, they suggested you gotta rip out all the code you have. You need to install this new library, and then it's gonna work. As CLI agent that you built. But I think Gemini suggested a very good integration that in just like three lines was able to solve my problem. And so I would say that a lot of times when it comes to coding agents, I think Gemini 3.5 Flash is a Incredibly capable model. We do have a pro model that's you know but we're testing and training and we're gonna ship when it's ready. A lot of times I think you just need to get used to the model. This is I think a model that we've been intensely dogfooding. So internally, tons of developers have been trying it. We've been getting feedback. And I think once we put now that we have the model out there, We're getting feedback from users, you know, we're putting it in frameworks like Mastra, we're trying things like OpenCode, we're trying to do with OpenClaw. As we get more as we get more feedback and we know what are the evaluations that You know we're not doing very well and you know if we're able to reproduce them. I'm very confident in the DeepMind team. I think we have one of the best research teams around. And so I think that a lot of it is just we want to basically get more feedback on people and how this model is doing and then it helps shape our roadmap, how we can train better models. Because this is a model that we've been intensely dogfooding ourselves.

Abhi: My initial reaction is 3.5 is really good at front end while GPT-5.5 is very bad. Those are the type of use cases you can start seeing Because you really have to get to know the model when you switch to a new one. I mean, I think everyone's been Opus-pilled for a long time, so they kinda don't know how to interact with other models. Even us internally, we're switching to 5.5 and people notice the temperament of the model changes immediately. What they expect in the dialogue changes. But we will be We'll be trying it out deeply and we'll give you some good feedback for sure.

Ivan: Yeah, for sure. Yeah. I think I would love to know how you guys are how you guys find it, where are some of the issues that, you know, it's not doing so well. Yeah, I think it helps us a lot in trying to shape our roadmap to see how we can train the next generation. Models. I mean I/O was just this intense sort of experience leading up to it. I think everyone was kind of working towards I/O as a goal and now that it's out I'm just It's just really exciting to see your products out there, you know. People are using the stuff that you ship and you're just like, oh, makes you feel very good about it.

Abhi: That's awesome So let's wrap it up on one last question I have for you. If you were to do a prediction for the rest of the year, what would you predict? In the AI community, AI space, what's What's one prediction? What's on your bingo card that you want to see checked off?

Ivan: I think we're gonna be rolling so many more agents in day-to-day life. I think we're just gonna see token consumption go up. I would love to see more people building agents Everyone's rolling different frameworks, the Mastra one, we have managed agents, but I think it's just gonna empower so many people to build automations and agents for day-to-day life. And so I think that's something I'm I feel like you're gonna see not just in the tech community, but in people that are just traditionally non-technical. My mom's asking me about OpenClaw now. And she's like, hey, what's this OpenClaw thing at? People keep following it to me Yeah, I'm like, oh, let me walk you through it. And I I think it's just it's really exciting to see a lot of people just start playing with models and feel the AGI, I guess. That to see a lot more people, it's just gonna be so normalized to just build and deploy so many more agents. Yeah.

Abhi: Well thanks Ivan. This is awesome. Thanks for walking us. Through everything. We'd love to have you on the show again in the future. Maybe after you know the next model release or any exciting things that you ever want to come and announce here. Floor is yours whenever you want.

Ivan: For sure. Thank you so much for having me.

Shane: And what's the best way for people to see what you're doing? Follow the DeepMind team. What what do you recommend?

Ivan: I'm on Twitter at Ivan Leo M K, I-V-A-N-L-E-O-M-K So that's probably the easiest way to reach out to me. I read every Twitter message that I can. And I think for us we we're very happy to hear feedback and how we can do better or how we're doing great.

Abhi: So yeah, everyone go give Ivan a follow and yeah We'll see you around, dude.

Ivan: Yeah, see you around, man.

Abhi: Take care.

Ad spot: Would you like to be a guest on the show? Visit mastra.ai/guest. Do you have a hot take? A problem you cracked? A cool product for a demo worth watching, share them, right here on Agent's Hour