⚡ Inside Google Labs: Building The Gemini Coding Agent — Jed Borovik, Jules + AIE CODE Preview

Okay. Jed Borovik, welcome to Lane Space.

Yeah, thanks for having me.

Uh, so we're sitting here at F.Ing's, uh, beautiful podcast studios, and we're actually meeting at GitHub Universe. Um, how's it been so far?

It's been great. Um, I mean, yeah, the keynote today was, was awesome. It was fun to see Jules up there a little bit, you know. But we have, uh, a lot of folks from our team here. Um, Jules is, yeah, partnering with, uh, GitHub for the new Agent HQ stuff, which we're excited about.

And also, this is a, this is an incredible podcast space, so yeah, I'm excited to-

Yeah

... do this here.

I'm glad for them to, to loan us this space. Uh, you are also an emcee for AI Engineer Code. Uh, that's exciting in New York, where you, you went to college, but you don't live there anymore.

Yeah. No, I spent a bunch of time in New York. You know, it's funny, um, being part of the New York tech scene, I actually think it's great having big major conferences there. Like, I think, uh, so there's a lot obviously that happens on the West Coast, um, but being someone in tech on the East Coast, it's, yeah, it's just awesome to have, have stuff there.

So I'm excited-

Yeah

... for it.

You mentioned you fly over to SF a lot, and like, like what's it... What, what's the scene like in the East Coast? Like, uh, like obviously we are pretty new and we're, like this is our first year coming to New York.

Um, what, what else happens in the New York-- Like, what are the highlights for you in the New York tech scene?

Yeah, I mean, there's so much. There's, uh, you know, obviously a ton of great companies. Um, I think the thing that's interesting about New York is it's such a big city with so much going on, right? And so there's like, you know, tech is a huge part of it, but there's also, you know, so many major industries there, whether it's-

Fashion

... media, fashion, finance. Like it's a, um... And so it's like, I think that helps push the tech, um, and do, do, do all kinds of stuff. Um, but yeah. No, the East Coast is a great city, great c- you know, the, the, all the schools, there's, you know, all across the East Coast, um, ton of great schools and great students doing all kinds of stuff.

Um, so yeah. You know, I went to school there. The hackathon scene there was amazing. Um, really fell in love with tech and, and programming there. So-

Is there a big NYU hackathon like, like, uh, in Stanford with like CalHacks and stuff?

Yeah. So there's, um-

But tree hacks.

Yeah, there's one that was put on by, this was a while ago, but put on by, uh, NYU and Columbia. We do a HackNY. Uh, so there's, there's, uh, there's a bunch of events kind of that, that we did together.

It would bring, you know, people across New York City, students across New York City, and those were super fun. Yeah, so it'd be at the Columbia one, one time and then NYU the next, and it would cycle back and forth.

So yeah, a lot of cool stuff was made there.

Nice. Uh, so you've been at Google for a while, nine years. Uh, you worked on a bunch of things, including with Malta, which, which is also another guest that I'm interviewing today. Uh, how'd you get into Jules? Like, what's the, what's the AI journey?

Yeah. So, you know, so this is gonna sound really cheesy, but I've told this story a couple times to folks when they're like, "Oh, how'd you end up, you know, doing this?" Um, but it is actually very true.

So I worked on Search for a long time, uh, and specifically kind of like news and freshness. Um, and then, uh, you know, when Stable Diffusion came out, that to me was the first like gen AI moment. I know some people talk about like ChatGPT as like the first thing, but for me, Stable Diffusion, uh, you know, it was a couple months before ChatGPT came out, it was a huge thing.

And I was, I was following it a ton online, and there were two groups of creators having reactions to it. You know, there was one group that was, um, you know, "This is stealing my art. This is stealing everything that's near and dear to me.

I hate this. This is ruining my life." And there was another group of artists and creators who were like, "Oh, this is a tool to create better art." And so I was watching this-

It's a new brush.

Yeah, exactly. And right around then, um, I was having conversations with a couple people who would say things like, "You know, if I had a kid in college, I wouldn't recommend they study computer science." I was like, "What?

Why?" And this was, you know, long before like Jensen Huang and people like, he had been saying this kind of stuff. I was like, "Whoa, whoa, why?" And it was like, "Oh, AI, like software engineering is gonna change.

It's gonna be so, you know, who knows if there are gonna be jobs." And I was like, "I love being a software engineer. I love programming," like... And I was like, "Wait, this is my Stable Diffusion moment. This is either it's gonna take my, my art, my craft, or this is a tool to create better art."

And I was like, "I definitely know which path I'm taking." So I got, you know, very into to building coding agents. So I was still working on Search, um, but I spent a bunch of time, you know, making stuff in my own time and playing with things and, um, and, you know, ultimately tried to find a role that would, you know, the most exciting role I could find to, to do this stuff.

And, and that was to join, uh, Google Labs and Jules, where we were, you know, right around then we were starting to build these kind of coding agents at Google. Um, and yeah, the timing worked out well, and, uh, I joined, and yeah, it's been, it's been awesome.

Can you, uh, well, since we're talking about Google Labs-

Yeah

... I am actually unclear about where Google Labs starts and the rest of, and then, and DeepMind and the rest of Google. Like, what, what is the org chart layer?

Yeah, yeah, yeah. That's a great question. Um, so Labs' mission is to build kind of new, like innovative products that the rest of Google isn't well-positioned for.

Yeah. Which, which we've had like, uh, the rise of from Opal GLM.

Yeah. Exactly, exactly.

So-

So Opal becomes maybe the, the, what most wide-

The most, yeah.

That's kind of-

And then NanoBanana, I don't know if it's...

Yeah. So, um, some of the-- So the thing about, that's really exciting about Labs is we work incredibly closely with DeepMind.

Yeah.

Right. So all the stuff in terms of the, you know, the we're building a product, but we work so closely for the model. And, you know, one of the nice things about being at Google is you have this opportunity to really build an end-to-end AI product, right?

From like pixels on a page, through the infrastructure, through, you know, the model and the training and all of that loop. So Labs is here to build new products, and we're really like a product org, but a true AI product org, where we work incredibly closely with, with, you know, uh, DeepMind, but also, you know, other parts of Google, uh, you know, as it makes sense.

Um-

Yeah. Just on the history of AI coding, I had heard that actually Google had an internal version of Copilot or something like that that was never released. Is that, is that true? What can we say about it?

Yeah. So, you know, I think there are, there are-- You know, Google's published papers in this space, uh, for a while. Um, and so yeah, we have, you know, being Google, we, we built a lot of our own tools and, you know, Cider, which, you know, folks maybe have heard of is our, you know, our internal IDE, and we've had all kinds of, you know, capabilities and tools there for a while.

So yes, you know, we certainly have had pretty good tools for a while, but they were for internal use.

Yeah. Yeah, I think, I think it was in-interesting because like I think the- Uh, one of the hype moments when Google started getting into the sort of, uh, like, LM game, like basically when everything rebranded to become Gemini and, like, starting to sort of push out Gemini, where people were like, "Oh, like, did you know that Google...

Probably, like, Google's entire repo is, uh, probably about the same size as GitHub." And like, I, you know, there, there must be some interesting data in there.

Oh, yeah. I mean, uh, and that's one of the things, you know, in building a lot of these internal systems is, you know, the, the, the data is incredible.

Yeah.

Especially when it's, you know, not only is the model and the training in-house, but all the data around the usage and whatever, so you know, we could build really kind of sophisticated, sophisticated things there.

Yeah. Okay. So let, let's, let's introduce people to Jules.

Yeah.

Uh, on the, your, your website says Jules autonomous coding agent. Uh, we've seen lots of these.

Yeah.

They, they're not octopuses. They're not purple. So you got that- You got that going for you. But, but, like, what, what really is, like, the core thing you're trying to nail in a very crowded coding agents landscape?

Yeah. So what we think about and what we set out to do, you know, back when I joined, was, like, where is, where are coding agents gonna go? And as these models get more and more powerful and sophisticated and, um, what is that experience gonna be?

And let's build for that future, right? And so when you think of a really powerful agent that can run for a really long time doing really complicated things, that's when, like, the products start to shape, take shape for us.

So for example, um, autonomous, you know, means, like, it has its own computer, right? So for Jules, it's, you know-

The VM.

The... Yeah, exactly. So tons of agents that run, you know, locally or in your workspace with you while you're coding. Um, but if you want something that's gonna run for hours or let's say days, you know, you might want it to have its own environment where it can, it can do its own work.

Um, so that's just one of the pieces that, that's imp- important for kind of this, uh, autonomous coding agent, but it's really like, you know, think about th- this future where they're incredibly powerful. You can spin up tons of them, right?

They're, they're autonomous, but also, you know, we've- we're thinking about what does, what does it mean for it to be ambient, right? Like, it's, it's, um, kind of when it has its own infrastructure and its own computer and its own w- ways to interact with it, um, how does that start to change what it can do?

For example, like, we have an API. So people are using it for all kinds of things, triggering it from, you know, when something happens, then, you know, we, we saw an example where someone has... They're triggering Jules for, to do kind of all kinds of updates to their site, and then they have a GitHub action that is going to automatically merge Jules pull requests.

So it's just, like, all kinds of stuff is flowing, really kind of-

Mm-hmm

... changing how people are, are able to do stuff. Um-

And is that CLI related, just to close that loop?

Yeah, CLI. So s- we also have a CLI, which is-

Yeah

... you know, we wanna meet developers where they are, right?

Yeah.

And so part of the, you know, an API is, like, you can trigger it from anywhere, but also, you know, when you're working locally, like, you wanna be able to trigger stuff. So, um, we, um... Yep. So we have, we have the, the, the Jules CLI we launched a couple weeks ago, which lets you interact with it.

By the time this podcast comes out, we'll be in the, integrated with the Gemini CLI. Um-

That's what I was thinking, like, you have a number of CLIs. I'm not sure.

Exactly.

Okay.

So Gemini CLI, um, all kinds of places where we're gonna kind of mix and, uh, and be, be able to harness this power, right? 'Cause, you know, developers work in all kinds of spots, um, and so making it easy, uh, easy to, to, you know, have this autonomous, ambient, um, agent that can really do all kinds of work for you.

Yeah. What is your journey? Like, when, when you started out, like, did you find any assumptions that were quickly challenged, you know, when working with GenAI and coding agents in, in, in general? Like, uh, I guess you, you're maybe not too unfamiliar with it because Search uses a, a lot of, like, machine learned, like, black boxy type things-

Yeah

... including BERT-

Yeah

... uh, which, you know, was, was, was a major update a few years ago. Uh, yeah, so I mean, just, just f- fill us in. Like, what is your AI engineering journey?

Yeah, totally. So, you know, I think one of the things that keeps coming up is, like, the, the model makes such a difference. I mean, maybe it sounds obvious, but it's like the, the quality of the model really changes, like, what you're able to do and how, how you engineer around it.

So for example, when we started, this was, you know, with relatively early models of Gemini, we had the agent scaffolding around it was incredibly complex. I think one of the things we've seen is scaffolds get simpler and simpler over time as the models get better.

And in some ways, this, the scaffolding is almost a crutch for, for, uh, things the, the, the model struggles with. For example, like, you know, really complicated sub-agent systems. You know, we've, we've played with that. We've, we've experimented with that.

Can you give an example of, like, a kind of sub- sub-agent that you had to abandon?

Yeah. It was just basically, like, you have, um, you know, you give, you give Jules a coding task to do, and it's going to have different agents for, um, whether it's, um, you know, making a code edit or handling a subproblem or, you know, doing any kind of action with an integration or, you know, and having, like, full sub-agents for, for different parts of it.

Like a reviewer agent or even, like, people, sometimes people do these, like, different personas where you're like, you know, one of the things that cracked me up is like, you know, you're the product manager agent, and then you have the code reviewer agent-

Oh, exactly

... the agent, this view. We, we didn't go, you know, that far. Um, but I think a lot of these things aren't as in favor now. I mean, certainly people do, you know, they, like, I don't want to say the agent harness isn't sophisticated.

It certainly is. But, um, you know, as the models get better, like, less is more, um, especially as it comes to, like, being able to improve through whether it's machine learning or, um, just, you know, regular maintenance. Um, I think certainly we, we found that, you know, we're finding that less is more.

Um, I think that, you know, that we were talking about a little bit before we started recording, like the, um, like RAG, right? And, and like, you know-

Codebase indexing

... indexing and all that stuff. And, you know, it seems like, you know, not just for, for Jules, but kind of across, across the industry, um, that, like, agent-based search, right? Like it's, you know, maintaining embeddings is hard, but getting the chunking right is hard.

Like-

Mm-hmm

... in terms of, like, the, the black box aspect you mentioned, um, a lot of that is, is hard to improve upon and, and, um-

Yeah. I would even say it's m- maybe not even hard so much as that it will never be good.

Well, tell me more. Why do you say that? Never-

Because, like, a chunk that happens to capture the thing you're looking for, um, you know, will c- will, will fail to capture something else.

Mm.

And so if you only retrieve based on, like, y- your embeddings of a, a chunk, like, it, it's, it uses very arbitrary boundaries that are drawn, like, ho- with, like, some hope of, like, so the semantics being captured, but you could just throw attention at it.

Oh. Yeah.

And you can scale probably much better using grep. Like

Totally. Totally. So I think that's, you know, that's an example of, of, you know, and these, these harnesses, how, like, they're simplifying, you know?

Yeah. Well, also, I, I haven't abandoned it completely because one of the things that we were doing, uh, I don't know if you saw the cognition sweet grep, uh, work, was basically using semant- semantic search or, and chunks in, in with embeddings as a tool.

Yep.

Uh, but it, on the same level as the other tools like-

Totally

... grep and, and, uh, FileXS and, and Glop and whatever else, uh, other variants you have. Uh, so I think like that, yeah, I mean, that, that, that makes sense. Like, don't abandon it-

Sure

... just, just don't reify it into-

Exactly. No

... like, the only way to do things.

Exactly. Exactly. And to be clear, like, you know, this is an area of research we're doing tons of work on. Um-

Yeah

... and, you know, I actually expect, you know, uh, in the coming months, we'll, we'll be talking about some stuff we're doing here too, but it's, um, it's, yeah, it's, it's not the... I feel like when we started, it was like RAG.

It was like embedding-based RAG.

Yeah.

It was, like, the thing everyone did, and it's interesting to see how it's changed.

People ask me for, like, "Where are the go code embedding models?" And, you know, I pointed them to, like, a few, like, Chinese ones. There's some, like, Nomic was working on one. And then, like, we found we didn't need them .

Yeah. Exactly. Exactly.

Very bitter lesson. So, so, so, you know, I, I think, like, that, these, this, like, these are good things. Like, I think, like, when Jules came out, it was kind of a preview. Uh, I, I'm in the, like, the trusted testers group-

Yeah

... so I, I got to see a little bit. Um, and but now it feels like more of a real product.

Yeah.

What's that transition like? Uh, is there a p- process within Google Labs to promote things when you feel like there's, there's some traction?

Yeah, absolutely. So I think the Google Labs is not, you know, about just experiments, right? So, like, you know, NotebookLM, as we talked, it was like a, you know-

It feels now very serious

... this incredibly successful product.

When we see money.

That's really, yeah. Um, and for us, I- IO was kind of a, a little bit of a turning point. Um-

Yeah

... so in, in May, when we, we announced, uh, Jules, you know, it was, like, great reception following IO. Um, and that was, that was a real moment of us to, like, turn this into, you know, a very much a real thing.

I mean, we did something that we were, you know, we always intended to. It wasn't ever intended, you know, I didn't j- you know, talk about my journey. Like, I didn't... It was always the goal to build, like, a real product here.

Um, and but for us, that, that was kind of a very key moment, very key milestone for us. Um, and so yeah, now it's, you know, it's very much a real thing. You know, we're, um, as mentioning, talking before, like, you know, Jules and the, you know, being talked about on the, in the GitHub keynote.

Um, uh, it's, yeah, it's certainly here to stay. We're, we're, um, we're, we're excited to kind of keep building and expanding.

Awesome. Let's talk about just, like, coding and just in general, your, uh, your coming to MCD AIE Code Summit. Uh, is it's gonna be your first time at AIE and, and EMC. Uh, what do you wanna know ?

Yeah, yeah, yeah. Well, to sell me. Why, why would someone wanna come-

Oh, gosh

... yeah, this is, yeah. Well, let's turn it around.

Oh, boy, this is embarrassing. Um, so- ... so I mean, we, you know, fortunately, we're in our third year, fourth year now, and we have a bunch of, you know, prior art we can just point people to and say, "Look at our YouTube.

Do you like that? You'll like this."

There's some great talks. You know, I haven't been before, but I've watched the talks.

Yeah. Yeah.

There's a lot of good stuff.

Yeah. Uh, and I- I'm proud that it features content from all labs, and basically we are, like, the, uh, this is the pattern I've seen across my career in terms of, like, every industry needs its, its, like, focal gathering points to just, like, trade tips and stuff.

Um, so I've seen that in JavaScript. I've seen that in cloud native. Uh, I've seen that in data engineering.

Mm-hmm.

And I was like, probably AI engineering will need something like this. And then I also, the, the, the concurrent thread to this was I went to a bunch of the academic ML conferences, NeurIPS, ICML, ICLR, and a lot of them, like NeurIPS is 40 years old, and it hasn't really changed and is very focused on academics and PhD students.

Whereas I think really, you know, the, the, the transition in AI going from research to industry is that it, you, you gradu- gradually sh- see a shift, uh, unfortunately, less open source, less papers-

Yeah, yeah

... and more products and more, uh, startups and, and closed models and, and what have you. But these people still wanna share. People still wanna, uh, hire. They wanna promote, uh, their work. Um, so they need a place to, to do that.

You can always do that at your company conferences, obviously, IO.

Yeah.

And, like, GitHub has GitHub and Microsoft has Build and Al- and Ignite. But, like, there usually is one place where it's, like, the industry neutral thing where everyone is on the same playing field, and may the best person win.

Yeah.

And, like, honestly, some people like that. Uh, you know, it's not like you're not gonna be treated as, like, the VIP and, like, you know, you kind of have to, like, earn your spot. But, like, when you earn your spot, I, I think, like, people give that, uh, that requisite level of attention better-

Yeah

... uh, bec- because you had to.

Yeah, yeah. Cool. So, you know, let's say I've watched, I've watched the videos online. I kind of get a sense for the speaker, but what's happening between that for someone who hasn't been before? Like, what goes on other than the talks?

Like-

Oh.

Yeah.

Yeah. Uh, a lot of, uh, well, just logistical stuff of, like, invoicing and, like, vendor selection and venue selection, and, like, did you know we have, like, five different pieces of software to, like, coordinate-

Okay

... speaker logistics and-

Yeah

... booth logistics and-

Yeah. But what does an at-

... AV

... an attendee, so I'm gonna go, I'm gonna sit-

Oh, yeah. Sorry

... in my house. But, uh, yeah. What am I gonna, what am I gonna get?

Yeah, yeah. So actually, uh, it's really weird because, like, as I'm the content guy for AIE, right?

Yeah.

I, I curate the speakers. I invite them. Right. And, uh, but I actually know that the content is, like, the least important part-

Yeah

... because all of it's filmed, and we're gonna edit it and post it for free on YouTube anyway. But the reason you come is because y- one, you can talk to the speakers, but also you can talk to each other.

Mm.

And so, like, the, the, you know, I always say, like, the hallway track is the most important track.

Yeah.

What's-

And how do you get the most out of the hallway track? What's your guide beginning a hallway track?

Oh, I don't have as collected a thought as, as I should. One, I think if you have some prior history of, like, what you're interested in and work on. So basically, like, the best intro to somebody is if they've seen you online before-

Okay

... so they can skip the whole, like, who the hell are you-

Right

... part and just get into like, "Oh, hey, I saw you wrote that thing. Like, let me talk to you in person about it since we're both here." Um, that's way better than like, "Who are you? What do you do?"

Sure.

And, and that's, and that's, like, a very cold interaction. Ideally, people come warm, or they can come with some clear idea of like, "Here's, here's why I'm here. Here's, here's what I'm looking to get out of this."

Mm-hmm.

Uh, because if, I think if you show up with, like, no, uh, real intention, or if you're, like, in and out for, uh, for your thing and nothing else, then you don't have the space and the mental f- energy for the unstructured, serendipitous connections.

And the thing about IE, at least in, at least in our scale, our size right now, and especially for the summits, which, which is the one that you're going to-

Mm

... um, everyone had to apply to get in.

Yeah.

Uh, so usually, uh, you know, our, our first, um, summit, we had like something like a 10 to one applicant to, to invite ratio, invited spots ratio.

Yeah.

This one's gonna be... And when it went up to, like, 10 to 16 to 20-something. This one's gonna be 23. Uh-

So one out of 23 people who applied get in?

Yeah, yeah. So, so, like, uh, yeah, it's, it's, it's a lot. Um, I think, like, um, it, it... and, and really we're trying to filter for people who would be speakers at any other conference. But, like, they, they are, they are at the top of the, the field.

They're either founders or honestly enterprise buyers-

Yeah

... of the, the best companies you can find in New York. Uh, which, you know... And, and that's another reason for this, the, our New York conference, which is we're bringing kind of the best of San Francisco or, or tech-

Mm-hmm

... to the, the, the finance sector, really.

Yeah.

Uh, there, there is a little bit of media, but mostly finance.

Yeah.

And, like, yeah, that's, that's great. Like, I mean, I, I think... So what I'm trying to say, I guess, is you're there to meet the other people, so make time to meet them. Have a calling card. Like-

Mm-hmm

... who are you? Like a, like a quick, like, what, who are you? What do you do? What, what can you help with? What are you looking for help for?

Yeah.

That kind of intro stuff is really good. Going with friends is really good. Obviously, like, we, we actually offer-- We, uh, for the world tour, we offer bundle discounts. This one I don't think we do. Uh, but just reach out if you, if you need something.

Uh but yeah, I mean, like, uh, I, I think, like, the idea of getting immersed in the code agent community is really important. Uh, we-- And I think maybe the, the last point I'll bring up is that we themed it for the first time, right?

So it used to be these are just generalists. Here's the state of AI, the best speakers we can get at any, any point in time. But now we're, we're really trying to push ourselves to theme everything, so we have the best people in code, the best people in data sets, the best people in RL.

I wanna do a mech interp one.

Mm.

That'll be fun.

Cool.

Uh, that one, that one I, I'm thinking it will be in London because, um, the, the people I wanna target are in London.

Yeah.

But yeah. I, I think, like, when you do a summit, it should be focused.

Mm-hmm.

Everyone there should have an agenda of, like, trying to learn what's the, what's the state-of-the-art, trying to have off-the-record conversations with their peers-

Mm

... doing the same thing at the other companies, and who knows what could happen, right? That's the, that's the weirdest thing. Like, I organize the thing, and I, I don't even know half the things that go on, just because my job is to provide the nexus-

Yeah

... of people to just connect.

Yeah.

Last time we were in New York, there were 13, maybe 15 side events organized by people, just, like, dinners, meetups, whatever, around the, around the summit, and we encourage it. We, we posted, uh, and we just want people to meet up.

Yeah. I was gonna ask, is there a whole, like, off-menu set of events happening? Like, how do people-

Yeah

... know?

They, uh, they, they organize it. Honestly, if you're, if you're not scared of strangers, you should organize your own.

Yeah.

Like a little dinner. Uh, we, we leave all the evenings open.

Okay.

So just, like, organize a dinner or a meetup focused on your thing. Like, we have people doing only voice.

Mm.

So, you know, if you wanna do voice, great. Uh, if you wanna do s- like, code review agents as, as a, as a small subset of generalist coding agents, do that. And, uh, I think you'll find it, right?

Like, or you can do, like, AI in finance, AI in bio, whatever, whatever the, the s- particular sector might be. Um, and I think, like, that is honestly the highest signal way to get, uh, a bunch of people who really resonate with your thing to, uh, to meet and, and have h- like, high bandwidth conversations.

Yeah, yeah. Are you maybe gonna do the autonomous, uh, coding agent dinner?

Uh, well, no. Uh, my job is to float.

He's messing around.

Yeah, yeah. My job is to handshake, ask, ask how everyone's doing, uh, fight fires. So I, I tend to just leave myself open, open until, you know, the end. But yeah. The... it'll, it'll be a, it'll be a sprint.

It's, it's, it's, uh, always a, a mad rush. And because I-- then I have to do my own talk. Uh, and-

There you go

... uh-

Do a bridge

... I totally screw me up. I don't know yet. I, I think, um, so far-- So, like, the last time I did this summit, I was talking about how this year had to, like, develop as the year of agents.

Mm.

And, like, it's really played out a lot.

Yeah.

Obviously, now, you know, the trendy thing is to say it's... no, it's not just a year, it's a decade of agents. But, like, um, this year, I think agents really took off, and most people got it right. Like, the consensus was correct.

You don't have to be too spicy or counter-consensus to say, like, if you worked on an agent, you're probably a lot better off. You probably made a lot of progress this year. Um, may- and maybe you can tell me how it feels from the Jules p-point of things.

I didn't see myself at the start of this year joining an agent company.

Mm.

And I ended up doing that. And but, like, I, I've gone so agent-pill to the point where, like-

Mm

... people come to me with startup ideas for infra companies. They're like, "What if we made a agent framework so that other people can build agents?" And I'm like, "Well, don't you just build agents yourself, bro?" Like-

There are a lot of- ... these frameworks, yeah.

Frameworks and infra companies.

Totally.

And all of these guys are just like, they're good developers with no conviction whatsoever in what they wanna build. They don't know what they-- what customer they want. They're just like- We wanna build developer tools so that's where we feel comfort-comfortable.

Yeah.

But honestly, it's not that hard to actually take a stand and be full stack and verticalize in some particular agent field that you want. Because guess what? They, like, the, the business and the economics are, are, you know, aligned that way.

Yeah.

And I'm not saying that you cannot make it as an infra company. There are some fantastic infra companies that, uh, are, are sponsors and, like, that I, I admire and, you know, I, I would invest in myself. It's just that comparatively, those are a lot harder .

Totally.

And, like, agent companies seem like they're shooting fish in a barrel. They seem like they're ramping up in ARR a lot faster, and it seem like their margins are better, so why not?

Yeah. So I mean, I think for us it's really been, "You're the agent." Like, as the models... You know, as we were talking about, what, what is-- Let's build Jules for where things are going. And as the models get better, I think it just becomes clearer and clearer that agents are super powerful.

You know, like we have, um, uh... You were talking about, like, before, high context and m-management, all that stuff's important. Like, we have people, we had-- This is a v- a funny story. We, we store some data for a session, but it only lasts-- We, we only store it for 30 days.

And so after 30 days, uh, your s- your session becomes locked. And when the first user starts, first started hitting that, they were upset. We were like, "There's no way anyone's gonna be using a single session for 30 days."

Like-

Yeah

... nobody would do a single track of work for 30 days. Uh, but just like how powerful that could be. So yeah, it's-

And how do you compress context when you run into it?

Yeah. So we have-- I mean, I can't talk too much about it, but we, you know, we do a lot of the standard things, and there's also, um, s- you know, we're developing a bunch of, of stuff. It's an active area of research for us.

Yeah. I, I think, like, uh, you know, just, just to... I, I'm not, not asking you for, for how exactly Jules does it.

Yeah.

There's just a number of approaches, right? And you just have to pick one.

Yeah.

Because you can't just s- use up your two million token context window. Is it, is it two million? Uh, it is up to two million.

Yeah. Especially for coding agents, 'cause like-

Yeah

... like you're reading files, like it's so... You're, you're running commands with huge outputs. Like, you know, I think coding agents are a really interesting area, both product-wise and the impact they're having, but also for research.

Yeah.

Like, they really push the limits of, you know, wha- what other domains are you running an agent for 30 days? What other domains, um, are you queuing so much context and so many turns and, um... So it's, uh, yeah, it's...

Coding agents are, I think, a kind of a special spot of, like, super interesting product, impact, research.

Yeah.

Yeah.

I see, I see the AMP folks drop the auto compaction-

Mm-hmm

... for a handoff mechanic-

Yeah

... which was pioneered by the Agents SDK.

Yeah.

Which is basically the sub-agents pattern where, like-

Yep

... you spin up a sub-agent and do a, do a thing. You don't need all their context that a sub-agent's doing, and then you, you can sort of come back to the, the main thread.

Totally. Yep. Yep, yep, yep. So it's... Yeah, it's a, uh, a good pattern. Uh, it also has its challenges, like how do you make sure you don't have stuff, you know, information is going back and forth, but that's the pattern.

You know, the, the summarization's a pattern. Um, you know, like, kind of s-externalizing some of that context, whether it's, like, writing it to, you know, like a note-

Yeah

... kind of thing is, is a common pattern. So yeah, there's, there's tons of things, uh, tons of things to try and do.

Yeah, yeah. I mean, and one thing I w- I do want to g-get more consensus about is what is the best, because I don't think I've read any papers-

Yeah

... about which, uh, methods compare better.

Yeah.

Um-

It's also interesting, like, as the models change, like, the answers change a little bit too.

Yeah, yeah. Claude, uh, you probably know, Claude externalizes too much.

Yeah. Yep. Yep, yep, yep.

Uh, how much does your work actually like... Uh, I feel like I, I switched back to, to Jules mode.

Yeah, yeah.

Uh-

Keep f-free flowing here.

But yeah, well, I mean, like, you know, how much does your work inform the model creation, right? Like, at the end of the day, like, you obviously are a very big consumer of Gemini models.

Yeah.

But also, you are not the only consumer, and they have other priorities than you.

Yeah, totally. Totally. I mean, I think we're lucky in kinda how we're positioned. We, we have very close relationships with, with DeepMind, so we have, um... And you know, coding agents are an important area. Like, let's be honest, right?

Like, for any kind of company building models, like, you can see it in all the labs, like, coding agents are important. Coding capabilities are really important.

Yeah, my, my, uh, OG image of the AIU code, um, I, I wrote something obnoxious like, "Code is the first spark of AGI."

Yeah, yeah.

Which is, like, probably true.

Totally. Uh, yeah. Uh, it's important from a kind of a AGI perspective. It's important from a dollars perspective. It's important for all of it. So it's, um... I think we're in a really lucky position where-

Yeah, yeah

... we have, uh, we're able to have a lot of kind of good collaboration and-

Yeah

... b-both ways, you know, like, all kinds of capabilities that are being developed and-

Yeah

... um... And you know what's interesting is it's, it's a whole host of things, right? 'Cause, you know, in terms of, like, AGI and the capability of these things, it's also, like, computer use models and browser use models, and so it's, it's, it's a, you know, models that output code.

But it's also the whole suite of, you know, things that you'd want an intelligent agent to be able to do. Um, so it's, uh, you know, multimodal. You know, it's all kinds of stuff, um, uh, that goes into it.

So it's, yeah.

What would you wanna find out from your peers at other coding agent companies? 'Cause you're, you're gonna meet all of them, basically.

Yeah. So I think one thing, and, you know, I, I don't think of this as a zero-sum thing. I think this is, like, really, like, uh, there's this tide that's gonna lift all of our boats, and, um, it's...

We're inventing a new way to do our art, right? You know, um, and how to create good art as a software engineer. And so what does that look like, and how does that f-feel? What does that, you know, what is the experience we wanna create?

I think as, as people working in AI, sometimes we don't do a good enough job describing this beautiful future we're creating. I mean, I know, like, you know, like, the CEOs and heads of these labs have started, like, you know, writing their, their think pieces on this, but- ...

right? Um, you know, for software engineers, like, what is this beautiful future we're creating? And like, you know, I think there's like... One, it's, it's inspiring. It makes it, you know, maybe less scary for, for people who are, who are thinking about these tools.

But also, like, you know, if, if we can't articulate it and think about it, it's less likely we'll get there, right? So, like, what is this, you know, great place we wanna create? Like, writing software is so hard.

Like, at so many companies, it's such a... And especially, like, big companies, it becomes so challenging to manage a code base and create. Um, and, and what can we do to make- You know, being a software engineer, inc- absolutely incredible experience.

What are these, you know, how do you want to interact with your model? How do you, how are you doing things locally versus in the cloud, and how does that interop? And, um, so I think like as an, as an industry, we're trying to like, you know, which is change.

Like we're, we're invi- in some ways inventing, and there's this movement to, you know, change how we do our art. Um, and yeah, the, the more, you know, the better we can create this experience, like we all, we, we all win to some degree.

Um, so, uh, yeah, I think that'd be one thing where it's like, yeah.

Yeah. The local to cloud sync is, um, the most contentious or important, I guess, topic for a lot of people. I wonder if we'll ever get like some kind of interop thing. Probably not, but, uh, a m-man can dream.

Tell me more about what, what's your dream, what's your dream flow here?

I don't know. Uh, start with Jules CLI, end up in Devon. I don't know.

Oh, interop between agents. Name these guys.

It's probably, it's probably meaningless, so this w- no. But like, I, I'm not actually serious about it. But like I-

Traffic to me all centric.

Yeah. Exist- Well, so, uh, I think Codex or is it Cloud Code? Cloud Code Web-

Mm-hmm

... can do this teleport-

Yep

... where they just basically dump like the entire history, and you can, you can pick it up in Cloud Code on, on your desktop. And probably that's the right move.

Yeah.

May- maybe there's, there's some more sort of elegant things, but they were first, so like, why not?

Yeah.

Uh, a- and like, and then, and actually maybe the, the, maybe the real thing is maybe it's not the conversation. M- maybe you don't need to teleport if the unit of, if the artifact that you pass back and forth is the linear ticket or the GitHub PR.

Mm-hmm.

Right? So you don't need the full JSON. Uh, you don't need the full chat history. You just need to pick up where other people left off because that's how humans do it.

Right, right, right.

Like I don't, I don't transfer my brain state to you. I just tell you what it did.

Yeah.

And then, you know, if I didn't, if I forgot to say something, you find out eventually.

Right, right. You say like the cloud agent, like dumps some kind of summary onto the ticket or whatever kind of it needs to-

Yeah

... pass on to the next-

In Slack or-

Yeah

... Linear and whatever.

Yeah. Yeah. That's interesting. I think we have s- there, there are some patterns emerging, the like IDE, CLI, cloud, right? Like these-

Yeah

... these are the pieces-

VS Code extension.

Yeah, VS Code, yeah. Like whether it's, you know-

Which you guys don't have yet maybe

... VC, VS Code. Like, like s- there's a, like the, the surface area is like standardizing it feels a little bit, um, and, um, yeah, how these things interop, how you can kind of make this like great experience between all of those.

Yeah, I think it's really interesting.

Yeah. Yeah, I think like... A- and then the other point, I just wanna backtrack a little bit to something else you said, which is like what the, the thick pieces that, uh-

Yeah

... these CEOs and stuff do. Uh, I, I think there's a lot of question about the impact that co- coding has on the software engineer industry in general, the humans.

Mm-hmm.

Do we end up-- do we stop hiring juniors altogether? Do we, uh, is it actually increasing productivity, or do you just feel like you're increasing productivity? I don't know if you have any take on that stuff.

Yeah, it's only so. I mean, I-- something we spend a lot, I spend a lot of time talking and thinking about w- with folks. And, um, we also, I also spend time talking to people at companies and, you know, I think sometimes working on these tools, it's interesting to see, uh, it's not as like diffused, this technology isn't as diffused across software engineers as I sometimes expect, right?

There's plenty of places that I think are, are not really using AI a ton. Uh, a lot of companies, a lot of software engineers aren't. Um, that being said, I'm very kind of excited about what this, what the future of software engineers are.

Like, could you imagine going back to not having these tools?

No.

That sounds horrible, right? Like that, um, and so, so th- that's one aspect of it. I also think, um, you know, I don't really buy this, like, you know, that we're not gonna hire more software engineers story. I think like for, for a few reasons.

Um, I mean, this is an example that, that often comes up, but is it like kind of the elasticity of the demand for software?

Okay. Yeah, Jevons paradox.

Exactly. And you know, like the, a lot of the cases sometimes come up is you look at like farming, right? And so, you know, there was a time in America where like the vast, vast majority of Americans were farmers, right?

And then technology happens, and today it's like less than one percent.

Yeah.

Um, and that's one example. But the flip side of that is you, you have electricity, which like, as that gets cheaper and cheaper, people just consume more and more and more electricity. Um, and you know, with food, there's only so much food we're gonna eat, right?

There's, there's a kind of a, a, only there's an inelastic demand for that, whereas electricity, very elastic demand. It seems like software, you know, software keeps getting better and better. Like the ability, like we're creating more and more software from like, you know, obviously like punch cards through to where we are today is like remarkably different in terms of how you're able to create software.

Um, so much more software is being made, and software just keeps becoming more and more of our GDP, right? Like it's, it's a, um... So I'm, I'm, I'm bullish on kind of the, the amount of software we'll be able to create, how it'll be created.

I think there's also something here about, you know, as a, as an engineer, being able to be more productive, like encourages more i- investment in s- people building software, right? If it's, you know, the job of a software engineer can now, you know, they can do 50% more, 100% more, 10X more, um, like justifying investment dollars into projects like dramatically changes, right?

And so, um, uh, yeah, I'm, I'm, I'm bullish on this idea that, that this is actually gonna be great for software engineers, both for, you know, our ability to kinda do our craft, our art, but also just what it means for the number of companies and the amount that's made and the quality of it and what we're able to do with it.

So, uh, yeah.

Yeah.

That's my rose-colored glasses take.

Rose-colored glasses indeed. Um, yeah, I have this take on the different kinds of work. Like we're s- we're splitting up the different kinds of software work, and there's a lot of commoditized work that we used to spend a lot of time on, and now we can basically entirely delegate to agents.

Mm-hmm.

And then that leaves us ideally for more strategic, important, novel, high-risk, whatever, uh, work, deep focus work that, uh, you know, is, is, is something I, I, I posted here on the, uh, semi-async value of death.

Yeah.

Where basically you kinda need to, uh, on the, in the extreme end, you can delegate to async agents-

Mm-hmm

... which Jules, uh, you know, uh, Cloud Code, whatever. But then over here, you kinda need the sort of- Deep involvement in understanding the code base and like-

Mm-hmm, mm-hmm

... feel like not vibe coding.

Totally.

Whatever the opposite of it is. Actually, that's my talk, which is... I've been thinking about this.

Okay.

So I tweeted out like, uh, this phrase 'cause I, I think, I feel it's in the air that like the term vibe coding was obviously coined by Andrej, and he's super influential, in February, and like people have just come to kind of use it as a blank check to just YOLO on prompts and stuff-

Yeah

... and, and create the worst code imaginable and like-

Yeah

... leave other people to clean it up.

Yeah.

Uh, so I think like people are kind of at their limits with this.

Mm-hmm.

Like, it was probably maxed out in terms of popularity. They would... And but we don't have yet what's next.

Right.

So my talk is really challenging every attendee, every speaker, to come up with like what is the aspirational good version of vibe coding that we can actually trust?

Yeah, what is it?

Well, like-

It's the punchline right now. What is it?

I, I mean, uh, the current leading candidate is agentic coding.

Mm-hmm.

Which is what Dharmesh Shah, who's like, I don't know if you know who Dharmesh is, uh, he's, he's pretty good track history when he's naming things.

Yeah.

Uh, it's just too many syllables. I don't think it just has the... It doesn't have the joy that-

Yeah

... that vibe coding invokes, which, uh, I think people want. But then, uh, people also want care and craft and like reliability and-

Yeah

... all that stuff that-

But if we don't have the term to describe it, maybe we don't have the pr- catchiest phrase for it, but what is, what is, what does it look like even if we don't have the phrase?

That, yeah, that's a great question. Um, I... Well, we have some speakers who are going to be pitching spectrum and development-

Mm-hmm

... that you have to really be thoughtful and effectively write a PRD. I think the... And, and I think like that is obviously very correct in terms of like, basically it's just a glorified prompt, but a very, very, very good one.

Mm-hmm.

And, uh, models are tuned to follow your prompt-

Yeah

... for good and for worse.

Yeah.

If you prompt sloppily, you're gonna get slop.

Yeah.

So a spec sounds good, I think. Uh, I don't know how often it'll be followed in practice-

Right

... because effectively what that transitions us to is a waterfall development approach where you spend three days writing a, writing a 50-page document, and then you kick off the agent. That doesn't seem right. Uh, so like, um, you know, obviously I, I, I have some bias here because, uh, cognition has from the start believed in interactive planning where like you kick off a thing, you get some feedback, y- then you're like, you're like, "Oh, that's not what I meant.

Let me correct myself."

Yeah.

'Cause I don't know what I wanted when I, when I started. So you, you work with the a- the, the machine to discover what you wanted, and the ma-ma-machine works with you to either get you what you wanted or show you the errors of your ways and, and then you correct it from there.

Yeah. I mean, one thing we talk about which very aligned is what you're thinking is like there are kind of like two problems as these things get better. One is like how do you specify what you want, and the other one is how do you verify that what you got is what you were thinking.

Yeah. Yeah.

And so, um, yeah, whether it's, you know, specifying through a spec or this like, you know, interactive plan or whatever it is. But then, yeah. And then on the flip side with the vibe coding thing is you might specify, but you never come back and verify, right?

Mm-hmm.

You're like you're... It's more hands off the wheel like maybe I'll click around the app a little bit and see how it works.

Mm-hmm.

But it's, I'm not really engaged with the code. So how do you, yeah, how are you verifying and making sure that it's, you know?

To my knowledge, you guys don't emphasize tests that much, right? It's not like you volunteer to write my tests.

Yeah. I mean, it depe- it depends. Like we, um, if there are tests in your code base, um, it's-

It's right. It's right out of the, the picture here. Jules will run your test suite.

Exactly. Exactly. So, um-

But it's not like, it's not like, you know, after everything, everything must have a matching test to the prompt that was mentioned, you know?

No.

That would be the extreme of-

Yeah

... what we mentioned.

I don't know if people always want that. I mean, maybe it'd be helpful to do that to kind of show that it was right, but let's say I don't write tests in my code base. Like, I want to merge that pull request that is introducing tests just for this one thing.

Like, you know, I think in so- in some ways the, the, the engineer should be able to control what, what kind of outputs they want.

Yeah.

Um, if it, if it helps and they want it, you know, absolutely. Um-

And then do you think there's other innovations on specifying apart from just chat?

Oh, totally. Totally. Um, I mean-

Agents, Agents MDE.

Yeah, agents, I mean, s- uh, Spec Development I think is in this, this category. I think, um, one area which is in, in is like multimodal.

Hmm.

Right? Like, you know, if, if I'm gonna show you a bug on our website, like, do I want to come and like type it with words to describe it or am I gonna point-

Yeah

... at the picture?

Yeah.

Um, and so, you know, with Jules you can upload images now. Um, but, you know, kind of more, you know, we have certain ways we communicate as humans that are easier in certain situations than others.

Yeah.

Let's, let's bring that to, to our engagement with, with the agents.

Yeah. So i- of all people, I expect you guys to be best at this because Gemini has video understanding. Just I want to submit a video-

Okay

... because some things I do cannot be screenshots.

Yep.

Uh, it's more about the behavior of, of things, uh, appear and disappear. Uh, yeah, I mean, I would love that if you, if you guys did it.

Yeah.

'Cause no, no one has it yet.

I know. I would love it too. We'll, we'll, I'll-

Oh, yeah

... I'll tag you and everyone else.

On my side, the visio- the, the version of that that we're exploring is Computer Use.

Yeah.

Uh, Computer Use was kind of introduced by Anthropic and then OpenAI did their totally with Operator and, and now Agent Mode in, in Atlas. I don't know if you guys have done anything super splashy on Computer Use, but anyway, it, it's coming back.

I, I can feel it.

Yeah. You know, I think, uh, yeah, definitely. And, you know, it, it ties into, you know, it ties into coding agents. It ties into just, you know, using AI systems in general.

Yeah. But basically, your VM now needs to render a UI or, or a browser, and then you need to let the agent click around in it.

Absolutely.

And you need to have precision-

Yep

... and speed and cost. The, and like, you know, affordable cost.

Yep.

It's a lot. It's, it's-

Yeah. No, these is... I mean, what can I sp- Prompts are so fun. There's just so much to build. There's so much... You know, I think also as a software engineer working in this space, like I think one of the reasons we, you know, you see so many companies in this space is partly like it's just so fun.

Like, there's so many things to build. There's so many tools, um, that seem like, you know, fun sci-fi. Like there's, it brings up a demo of what I've worked on. It's clicking around, and I can see a video of it, or I can even take over and use it like...

So, uh, yeah, it's, yeah.

Awesome. Okay. So just moving towards wrapping up.

Yeah.

If people run into you at AIE, uh, they've, you know, they heard your, your, your pitch on Jules-

Yeah

... what else should they also talk to you about? Like, you know, what, what, what are you, what can you help with versus what are you looking for?

Anyone should feel free to come up and talk to me, you know, at any point. I, uh, you know, obviously very interested in anyone who's doing stuff with coding agents or someone who's using coding agents in interesting ways.

I'm always curious about, you know, workflows people have with their coding agents, whereas, you know, whether it's, you know, "Hey, I'm using this tool in this way, and I've, you know, configured this crazy thing." Like, I always love hearing how people are using it.

I also love hearing people who are having bad times with it, where it's like, actually I don't, you know, maybe they're not coming to this conference, but, you know, I've, I've, "Jeff, tried all these co- tools, and I don't like them.

I don't use them, and here's why." Um-

Yeah

... so, you know, I'm totally open for any side of, of the, uh, all the way from, you know, full AI pilled and coding, coding AI lovers to people who hate it. As far as what I'm looking for, you know, I think, um, you know, really ju-just going to kind of connect and meet people.

I think, you know, we are always hiring, so like, you know, I'm, I, uh, anyone who's, you know, interested in working on this stuff, um, I'm always happy to talk. But yeah, really just kind of, you know, meeting people, spending time geeking out on this stuff.

Yeah. There'll be lots of geeking out.

Yeah.

Uh, all right. Thanks for your time. Looking forward.

Yeah, same.