⚡ Google Labs 内部：打造 Gemini Coding Agent——Jed Borovik、Jules 与 AIE CODE Preview

Okay, Jed Borovik, welcome to Latent Space.

Yeah, thanks for having me.

We're sitting here at Figma's beautiful podcast studios, and we're actually meeting at GitHub Universe. How's it been so far?

It's been great. The keynote today was awesome. It was fun to see Jules up there a little bit. We have a lot of folks from our team here. Jules is partnering with GitHub for the new Agent HQ stuff, which we're excited about, and this is an incredible podcast space. I'm excited to do this here.

I'm glad for them to loan us this space. You're also MC for AI Engineer Code. That's exciting in New York, where you went to college, but you don't live there anymore.

I spent a bunch of time in New York. It's funny being part of New York's tech scene. I actually think it's great having big, major conferences there. There's a lot that happens on the West Coast, but being someone in tech on the East Coast, it's awesome to have stuff there.

You mentioned you fly over to San Francisco a lot. What's the scene like on the East Coast? Obviously, we're pretty new—we're here for our first year coming to New York. What else happens in New York? What are the highlights for you in the New York tech scene?

There's so much. There's obviously a ton of great companies. I think the thing that's interesting about New York is that it's such a big city with so much going on. Tech is a huge part of it, but there are also so many major industries there, whether it's media, fashion, or finance. In some ways, I think that helps push the tech and lets people do all kinds of stuff.

The East Coast is a great place. All the schools are spread across the East Coast, with a ton of great schools and great students doing all kinds of stuff. I went to school there. The hackathon scene there was amazing. I really fell in love with tech and programming there.

Is there a big NYU hackathon, like Stanford with Cal Hacks and stuff?

Yeah.

TreeHacks?

Yeah. There's one that was put on by NYU and Columbia. We did it together—HackNYU. There were a bunch of events that we did together, bringing people across New York City—students across New York City—and those were super fun. It'd be the Columbia one one time, then NYU the next, and we'd cycle back and forth. A lot of cool stuff was made there.

Nice. You've been at Google for a while—9 years. You've worked on a bunch of things, including with Malte, who's also another guest I'm interviewing today. How did you get into Jules? What's the AI journey?

This is going to sound really cheesy, but I've told this story a couple of times to folks when they're like, “How did you end up doing this?” It is actually very true. I worked on Search for a long time, specifically news and freshness.

When Stable Diffusion came out, that was the first gen-AI moment for me. I know people talk about ChatGPT as the first thing, but for me, Stable Diffusion—it was a couple of months before ChatGPT came out—was a huge thing. I was following it a ton online, and there were 2 groups of creators having reactions to it.

There was one group of artists and creators saying, “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 saying, “This is a tool to create better art.” I was watching this bifurcation.

Yeah, exactly.

Right around then, I was having conversations with a couple of people who would say things like, “If I had a kid in college, I wouldn't recommend they study computer science.” I was like, “What? Why?” This was long before Jensen Huang and people like him had been saying this kind of stuff. I was like, “Whoa, why?”

They'd say, “Oh, AI. Software engineering is going to change. Who knows if there are going to be jobs?” I loved being a software engineer. I love programming, and I was like, “Wait, this is my Stable Diffusion moment. This is either going to take my art, my craft, or this is a tool to create better art.” I definitely knew which path I was taking.

So I got very into building and coding. I was still working on Search, but I spent a bunch of time making stuff in my own time and playing with things. Ultimately, I tried to find the most exciting role I could find to do this stuff, and that was to join Google Labs to work on Jules, where we were starting to build these kinds of coding agents at Google. The timing worked out well, and I joined. It's been awesome.

Since we're talking about Google Labs, I'm actually unclear about where Google Labs starts, and then the rest of Google, DeepMind, and the rest of Google. What's the org-chart layer?

Labs' mission is to build new, innovative products that the rest of Google isn't well-positioned for.

Yeah, which we've had, like, Riser from Nom[?].

Exactly. Exactly. So it's maybe the most widely known, and then Nano Banana. I don't know if it's—

Yeah. The thing that's really exciting about Labs is that you work incredibly closely with DeepMind.

Right. We're building a product, but we work so closely with DeepMind on the model. One of the nice things about being at Google is that you have the opportunity to really build an end-to-end AI product, from the pixels on the page through the infrastructure, the model, the training, and all of that loop.

Labs is here to build new products, and we're really a product organization, but a true AI product organization, where we work incredibly closely with DeepMind, as well as other parts of Google as it makes sense.

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

I think Google has published papers in this space for a while. In Google, we've built a lot of our own tools, and Cider, which folks may have heard of, is our internal IDE. We've had all kinds of capabilities and tools there for a while. So, yes, we've certainly had pretty good tools for a while, but they were for internal use.

I think it was interesting because one of the hype moments, when Google started getting into the LLM game—basically when everything rebranded to become Gemini and started pushing out Gemini—people were like, “Oh, did you know that Google's entire repo is probably about the same size as GitHub? There must be some interesting data in there.”

Oh, yeah. One of the things in building a lot of these internal systems is that the data is incredible, especially when not only is the model and training in-house, but all the data on usage and everything. We could build really sophisticated things there.

Okay, let's introduce people to Jules. Your website says, “Jules: autonomous coding agents.” We've seen lots of these. They're not octopuses. They're not purple, so you've got that going for you. What are you really trying to nail at the core in a very crowded coding-agent landscape?

What we think about, and what we set out to do back when I joined, was: where are coding agents going to go as these models get more and more powerful and sophisticated, and what is that experience going to be? Let's build for that future. When you think of a really powerful agent that can run for a really long time doing really complicated things, that's when the product started to take shape for us.

For example, autonomous means it has its own computer. For Jules, it's in the cloud, exactly. Tons of agents run locally or in your workspace with you while you're coding, but if you want something that's going to run for hours or, let's say, days, you might want it to have its own environment where it can do its own work.

That's one of the pieces that's important for this autonomous coding agent, but it's really about thinking about this future where they're incredibly powerful. You can spin up tons of them. They're autonomous, but we're also thinking about what it means for them to be ambient. When it has its own infrastructure and its own computer and its own ways to interact with it, how does that start to change what it can do? For example, we have an API, so people are using it for all kinds of things.

Triggering it when something happens—we saw an example where someone is triggering Jules to make all kinds of updates to their site. Then they have a GitHub Action that's going to automatically merge Jules's pull requests. All kinds of stuff is flowing, really changing how people are able to do things.

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

Yeah, CLI. We also have a CLI. We want to meet developers where they are, right? An API means you can trigger it from anywhere, but when you're working locally, you want to be able to trigger things, too. We launched the Jules CLI a couple of weeks ago, which lets you interact with it. By the time this podcast comes out, we'll be integrated with the Gemini CLI.

That's what I was thinking—you have a number of CLIs.

Exactly. So, Gemini CLI, all kinds of places where we're going to mix and where you'll be able to harness this power. Developers work in all kinds of places, so we're making it easy to have this autonomous, ambient agent that can do all kinds of work for you.

What was your journey like when you started? Did you find any assumptions that were quickly challenged when working with Gemini and coding agents in general? I guess you're maybe not too unfamiliar with it, because Search uses a lot of machine-learned, black-box-type systems, including BERT, which was a major update a few years ago. Fill us in: what is your AI engineering journey?

Totally. One of the things that keeps coming up is that the model makes such a difference. It may sound obvious, but the quality of the model really changes what you're able to do and how you engineer around it. When we started, we were working with relatively early models of Gemini, and the agent scaffolding around it was incredibly complex.

One of the things we've seen is that scaffolds get simpler and simpler over time as the models get better. In some ways, the scaffolding is almost a crutch for things the model struggles with. For example, we've played with and experimented with really complicated subagent systems.

Can you give an example of a subagent system that you had to abandon?

It was basically like this: you give Jules a coding task, and it's going to have different agents for making a code edit, handling a subproblem, or taking an action with an integration. You might have full subagents for different parts, like a reviewer agent. People sometimes use different personas, where you have a product manager agent and then a code reviewer agent. That cracked me up.

In this video, we didn't go that far. I think a lot of these things aren't as much in favor now. I don't want to say the agent harness isn't sophisticated—it certainly is—but as the models get better, less is more, especially when it comes to improving through machine learning or regular maintenance.

I think we were talking a little bit before we started recording about RAG, and mixing and matching all of that. It seems like, not just for Jules but across the industry, agent-based search is difficult. Maintaining embeddings is hard, but getting the chunking right is hard, too. In terms of the black-box aspect you mentioned, a lot of that is hard to improve upon.

I would even say it's maybe not so much hard as it is something that will never be good. [laughter]

Why do you say it will never be good?

A chunk that happens to capture the thing you're looking for will fail to capture something else. If you only retrieve based on the embeddings of a chunk, you're using very arbitrary boundaries that are drawn with some hope that the semantics will be captured. But you could just throw attention at it. [laughter]

Totally. You can probably scale much better using GPUs.

Totally. [laughter] So I think that's an example of how these harnesses are simplifying.

I haven't abandoned it completely, because one of the things we were doing—I don't know if you saw the Cognition's SWE work—was basically using semantic search, chunks, and embeddings as a tool, but on the same level as the other tools, like web and file access, grep, and whatever other variants you have. I think that makes sense: don't abandon it; just don't reify it into the only way to do things.

Exactly. And to be clear, this is an area of research we're doing tons of work on. I expect that in the coming months, we'll be talking about some of the stuff we're doing here, too. It's not the only way. I feel like when we started, it was embedding-based RAG. It was like the thing everyone did, and it's interesting to see how it's changed.

People asked me where the good code-embedding models were. I pointed them to a few Chinese ones. Nomic was working on one, and then we found we didn't need them.

Yeah. [laughter] Exactly. Exactly.

Very bitter lesson.

Yeah.

What's that transition like? Is there a process within Google Labs to promote things when you feel like there's some traction?

Absolutely. Google Labs isn't just about experiments. NotebookLM, for example—

It's not a very serious, incredibly successful product that makes money?

It's really not just that. For us, Jules was kind of a turning point. In May, when we announced Jules, it received a great reception following I/O. That was a real moment for us to turn this into a very real thing. It was something we always intended to build; it was always a goal to build a real product here.

That was a very key moment and milestone for us. Now it's very much a real thing. As we were talking about before, Jules was mentioned in the GitHub keynote. It's certainly here to stay, and we're excited to keep building and expanding.

Awesome. Let's talk about coding in general. You're coming to the MCP AI Engineer Code Summit. It'll be your first time at AI Engineer and MCP. What do you want to know?

Yeah. Well, tell me why someone would want to come. Let's turn it around.

Oh boy. This is embarrassing. Fortunately, we're in our third year—fourth year now—and we have a bunch of prior art we can point people to and say, “Look at our YouTube. Do you like that? Do you like this?”

There's some great talks. I haven't been before, but I've watched the talks. There's a lot of good stuff.

I'm proud that it features content from all labs. This is a pattern I've seen across my career: every industry needs focal gathering points to trade tips and share things. I've seen that in JavaScript, cloud-native, and data engineering. I figured AI engineering would need something like this, too.

The concurrent thread was that I went to a bunch of the academic ML conferences, like NeurIPS, and a lot of them are decades old. NeurIPS is 40 years old, and it hasn't really changed. It's very focused on academics and PhD students, whereas I think the transition in AI from research to industry means you gradually see a shift—unfortunately, less open source and fewer papers, and more products, startups, and closed models.

People still want to share. They still want to hire, and they want to promote their work, so they need a place to do that. You can always do that at your company conferences. Obviously, Google has I/O, GitHub has GitHub Universe, and Microsoft has Build and Ignite, but there usually is one industry-neutral place where everyone is on the same playing field.

Honestly, some people like that. It's not like you're going to be treated as a VIP; you kind of have to earn your spot. But when you earn your spot, I think people give it the requisite level of attention, because you had to.

Yeah. Of course. So, you know, I’ve watched the videos online. I kind of get a sense of the specifics, but what’s happening beyond that for someone who hasn’t been before? What goes on other than the talks?

Oh, yeah. A lot of logistical stuff: invoicing, vendor selection, venue selection. Did you know we have 5 different pieces of software to coordinate speaker logistics, room logistics, and attendee logistics? So I’m going to go sit [laughter], but—

Yeah, what am I going to get?

Yeah, yeah. So actually, it’s really weird because I’m the content guy for AI Engineer, right? I curate the speakers I invite. But I actually know that the content is the least important part, because all of it’s filmed and we’re going to edit it and post it for free on YouTube anyway.

But the reason you come is because, one, you can talk to the speakers, but also you can talk to each other. I always say the hallway track is the most important track.

Yeah. [laughter] How do you get the most out of the hallway track? What’s your guide to the hallway track?

I don’t have as collective a set of thoughts as I should. One, I think it helps if you have some prior history of what you’re interested in and what you work on. Basically, the best intro to somebody is if they’ve seen you online before, so they can skip the whole “Who the hell are you?” part and just get into, “Hey, I saw you wrote that thing. Let me talk to you in person about this, since you’re both here.” That’s way better than, “Who are you? What do you do?”

That’s a very cold interaction. Ideally, people come warm, or they can come with some clear idea of, “Here’s why I’m here. Here’s what I’m looking to get out of this.” Because if you show up with no real intention, or if you’re in and out for your thing and nothing else, then you don’t have the space and the mental energy for the unstructured, serendipitous connections.

The thing about—at least at our scale, our size right now, especially for the summits, which is the one that you’re going to—is that everyone had to apply to get in. Usually, our first summit had something like a 10-to-1 applicant-to-invited-spots ratio. This one went up to 16 to 20-something; this one’s going to be 23. So, 1 out of 23 people who apply gets in.

Yeah, it’s a lot. We’re really trying to filter for people who would be speakers at any other conference, but they are the top of the field. They are either founders or, honestly, enterprise buyers of the best companies you can find in New York. That’s another reason for our New York conference: we’re bringing the best of San Francisco tech to the finance sector, really.

There is a little bit of media, but mostly finance. What I’m trying to say, I guess, is that you’re there to meet the other people, so make time to meet them. Have a calling card: Who are you? What do you do? What can you help with? What are you looking for help with? That kind of intro stuff is really good.

Going with friends is really good. Obviously, we actually offer bundle discounts for the AI Engineer World’s Fair. This one, I don’t think we do, but just reach out if you need something.

The idea of getting immersed in the code agent community is really important. We themed it for the first time, too. These used to be generalist events: here’s the state of AI and the best speakers we can get at any point in time. But now we’re really trying to push ourselves to theme everything. We have the best people in code, the best people in datasets, and the best people in RL. I want to do a mech interp one.

That’ll be fun.

That one, I’m thinking, will be in London, because the people I want to target are in London. But, yeah, I think when you do a summit, it should be focused. Everyone there should have an agenda of trying to learn what the state of the art is and trying to have off-the-record conversations with their peers doing the same thing at other companies.

Who knows what could happen? That’s the weirdest thing. I organized a thing, and I don’t even know half the things that go on, just because my job is to provide the nexus of people to connect.

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

Yeah. I was going to ask: is there a whole off-main-stage set of events happening? How do people know?

Yeah, they organize it. Honestly, if you’re not scared of strangers, you should organize your own little dinner. We leave all the evenings open.

Okay.

Just organize a dinner or a meetup. Focus on your thing. We have people doing only voice, so if you want to do voice, great. If you want to do code review agents, as a small subset of generalist coding agents, do that.

I think you’ll find it, right? Or you can do AI in finance, AI in bio—whatever that particular sector might be. That is honestly the highest-signal way to get a bunch of people who really resonate with your thing to meet and have high-bandwidth conversations.

Yeah. Yeah. Are you and I going to do the autonomous coding agent dinner?

Well, no. [laughter] My job is to float. My job is to handshake, ask how everyone’s doing, and see if I vibe. I tend to just leave myself open until the end.

But, yeah, it’ll be a sprint. It’s always a mad rush because then I have to do my own talk, and I don’t know yet.

I think, so far, the last time I did this summit, I was talking about how this year had to develop into the year of agents. It’s really played out a lot. Obviously, now the trendy thing is to say, no, it’s not just the year; it’s the decade of agents.

But this year, I think agents really took off, and most people got it right. The consensus was correct. You don’t have to be too spicy or counter-consensus to say that if you worked on an agent, you’re probably a lot better off. You probably made a lot of progress this year.

Maybe you can tell me how it feels from the Jules point of view. I didn’t see myself at the start. You’re joining an agent company—

—and I ended up doing that. I’ve gone so agent-pilled to the point where people come to me with startup ideas for infrastructure companies. They’re like, “What if we made an agent framework so that other people couldn't build agents?” I’m like, “Why don’t you just build agents yourself, bro?”

Like, there are a lot of frameworks—frameworks and infrastructure companies. All these guys are just good developers with no conviction whatsoever in what they want to build. They don’t know what customer they want. They’re just like, “We want to build developer tools,” so that’s where we feel comfortable.

But honestly, it’s not that hard to take a stand and be full-stack and verticalize in some particular agent field that you want. Guess what? The business and the economics are aligned that way.

I’m not saying that you cannot make it as an infrastructure company. There are some fantastic infrastructure companies that I admire and would invest in myself. It’s just that, comparatively, those are a lot harder. 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 seems like their margins are better, so why not?

Yeah. So, I mean, for us, it’s certainly been— as the models get better, and as we were talking about, “What is it? Let’s build Jules for where things are going”—I think it just becomes clearer and clearer that agents are super powerful.

High-context management and all that stuff is important. We had a funny story: we store some data for a session, but it only lasts 30 days. We only store it for 30 days, and so after 30 days your session becomes locked. When the first user started hitting that, they were upset. We were like, “There’s no way anyone’s going to be using a single session for 30 days.” Are people doing a single track of work for 30 days? That’s just how powerful that could be.

How do you compress context when you run into it?

We have—I can’t talk too much about it, but we do a lot of the standard things. We’re also developing a bunch of stuff. It’s an active area of research for us.

Yeah. I think, just to be clear, I’m not asking you exactly how Jules does it. There are a number of approaches, right? You just have to pick one, because you can’t just use up your 2-million-token context window.

Is it 2 million? It is up to 2 million, especially for coding agents, because you’re reading files and running commands with huge outputs. I think coding agents are a really interesting area, both product-wise and in terms of the impact they’re having, but also for research. They really push the limits: what other domains are you running an agent for 30 days, and what other domains are you accumulating so much context and so many turns? Coding agents are kind of a special spot—super interesting from a product-impact and research perspective.

Yeah, I see the Anthropic folks have dropped auto-compaction for a handoff mechanic, which was pioneered by the Agents SDK.

Which is basically the subagents pattern: you spin up a subagent, have it do a thing, and you don’t need all of the context that subagent is using.

And then you can sort of come back to the main thread.

Totally. Yep, yep, yep. It’s a good pattern, but it also has this challenge: how do you make sure enough information is going back and forth? Summarization is a pattern. Externalizing some of that context, whether it’s writing it to a note or something like that, is a common pattern. There are tons of things to try and do.

Yeah. One thing I do want to get more consensus about is what the best method is, because I don’t think I’ve read any papers—

—about which methods compare better.

Yeah. As models change, the answers change a little bit too.

Yeah, yeah. Claude probably externalizes too much.

Yeah. [laughter]

Yeah.

How much does your work actually— I feel like I switched back to Jules mode.

Yeah, yeah. Keep it free-flowing here.

Yeah. Well, I mean, how much does your work inform model creation? At the end of the day, you’re obviously a very big consumer of Gemini models, but you’re not the only consumer, and they have other priorities than you.

Yeah, totally. Totally. I think we’re lucky in how we’re positioned. We have very close relationships with DeepMind, and coding agents are an important area. Let’s be honest: for any company building models, you can see it in all the labs—coding agents are important, and coding capabilities are really important.

My OG image of the AI coding UI—I wrote something obnoxious like, “Code is the first spark of AGI.”

Yeah.

Which is probably true.

Totally. Yeah, it’s important from an AGI perspective. It’s important from a dollars perspective. It’s important for all of it. I think we’re in a really lucky position where we’re able to have a lot of good collaboration, both ways. All kinds of capabilities are being developed, and it’s interesting because it’s a whole host of things. In terms of AGI and the capabilities of these systems, there are also computer-use models and browser-use models. It’s models that output code, but it’s also the whole suite of things that you’d want an intelligent agent to be able to do. It’s multimodal; it’s all kinds of stuff that goes into it.

What would you want to find out from your peers at other coding-agent companies? You’re going to meet all of them, basically.

Yeah. I think one thing—and I don’t think of this as a zero-sum thing—is that there’s a tide that’s going to lift all of our boats. We’re inventing a new way to do our art: how to create good art as a software engineer. What does that look like, and how does that feel? What is the experience we want to create?

As people working in AI, sometimes we don’t do a good enough job describing this beautiful future we’re creating. I know the CEOs and heads of these labs have started writing their think pieces on this, but for software engineers, what is this beautiful future we’re creating? I think that’s inspiring. It makes it less scary for people who are thinking about these tools, but if we can’t articulate it and think about it, it’s less likely we’ll get there.

What is this great place we want to create? Writing software is so hard. At so many companies, especially big companies, it becomes so challenging to manage a codebase. What can we do to make being a software engineer an absolutely incredible experience? How do you want to interact with your model? How are you doing things locally versus in the cloud, and how do they interoperate?

As an industry, we’re trying to change how we do our art. In some ways, we’re inventing it. The better we can create this experience, the more we all win to some degree.

Yeah, local-to-cloud sync is the most contentious—or important, I guess—topic for a lot of people. I wonder if we’ll ever get some kind of interrupt thing. Probably not. But in your dream, tell me more about your dream flow here.

I don’t know. Start with the Jules CLI, end up in Devin. I don’t know.

Oh, interaction between agents, yeah.

It’s probably meaningless. So, no, I’m not actually serious about it, but—

Actually, well, I think Codex—or is it Claude Code Web?—can do this teleport.

Mhm. It can do this teleport.

Yep.

They basically dump the entire history, and you can pick it up in Claude Code on your desktop. That’s probably the right move.

Yeah.

Maybe there are more elegant things, but they were first, so why not?

Yeah.

Maybe the real thing is that it’s not the conversation. Maybe you don’t need to teleport if the unit of the artifact that you pass back and forth is the Linear ticket or the GitHub PR.

Right. You don’t need the full JSON. 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.

Right. I don’t transfer my brain state to you. I just tell you what I did.

Yeah.

And then, if I forgot to say something, you find out eventually. The cloud agent dumps some kind of summary onto the ticket, or whatever it needs to pass on to the next—

—in Slack, Linear, and whatever.

Yeah, yeah. That’s interesting. There are some patterns emerging, though: IDE, CLI, cloud. Those are the pieces, along with a VS Code extension.

Yeah, VS Code.

Whether you guys have one or not, the surface area is standardizing. It feels a little bit—how these things interoperate and how you can make this a great experience across all of those is really interesting.

Yeah. I want to backtrack a little bit to something else you said: the think pieces these CEOs and others put out. I think there are a lot of questions about the impact that coding has on the software-engineering industry in general, on the humans. Do we stop hiring juniors altogether? 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, totally. This is something I spend a lot of time talking and thinking about with folks. I also spend time talking to people at companies, and sometimes, working on these tools, it’s interesting to see that this technology isn’t as diffused across software engineers as I sometimes expect. There are plenty of places that aren’t really using AI a ton. A lot of companies and a lot of software engineers aren’t.

That being said, I’m very excited about what the future of software users looks like. Could you imagine going back to not having these tools? No, that sounds horrible. That’s one aspect of it.

I also don’t really buy this story that we’re not going to hire more software engineers, for a few reasons. This is an example that often comes up: the elasticity of the demand for software.

Okay.

Jevons paradox.

Exactly.

A lot of the cases that sometimes come up involve farming. There was a time in America when the vast majority of Americans were farmers, and then technology happened. Today, it’s less than 1%.

Yeah.

That’s one example, but the flip side is electricity. As it gets cheaper and cheaper, people just consume more and more and more electricity. With food, there’s only so much food we’re going to eat. There’s an inelastic demand for that, whereas software has very elastic demand.

It seems like software keeps getting better and better. The ability to create more and more software, from punch cards through to where we are today, is remarkably different in terms of how you're able to create software. So much more software is being made, and software keeps becoming more and more of our GDP. I'm bullish on the amount of software we'll be able to create and how it will be created.

I think there's also something here about how, as an engineer, being able to be more productive encourages more investment in people building software. If a software engineer can now do 50% more, 100% more, or 10x more, justifying investment dollars into projects dramatically changes. I'm bullish on the idea that this is going to be great for software, both for our ability to do our craft and for what it means for the number of companies, the amount that's made, the quality of it, and what we're able to do with it. That's my rose-colored-glasses take.

Rose-colored glasses indeed. I have this take on the different kinds of work. 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 that we can now basically entirely delegate to agents.

Then that leaves us, ideally, for more strategic, important, novel, high-risk—whatever—work: deep, focused work that is meaningful.

I feel like it's not vibe coding—whatever the opposite of it is. Actually, that's my talk. I've been thinking about this.

So I tweeted out this phrase because I think it's in the air. The term “vibe coding” was obviously coined by Andrej Karpathy, who's super influential, in February, and people have come to use it as a blank check to just YOLO on prompts and create the worst code imaginable, leaving other people to clean it up.

I think people are reaching their limits with this. It was probably maxed out in terms of popularity, but we don't yet have what's next.

Right?

My talk is really challenging every attendee and every speaker to come up with the aspirational, good version of vibe coding that we can actually trust.

Yeah. What is it?

The punchline right now—the current leading candidate—is agentic coding, which is what DHH says. I don't know if you know who DHH is, but he has a pretty good track record when he's naming things.

It's too many syllables. I don't think it has the joy that vibe coding invokes, which I think people want. But people also want care, craft, reliability, and all that stuff.

If we don't have the term to describe it, maybe we don't have a catchy phrase for it, but what does it look like? Even if we don't have the phrase, what does it look like?

That's a great question. We have some speakers who are going to be pitching spec-driven development, where you have to be really thoughtful and effectively write a PRD. I think that is obviously very correct in the sense that it's just a glorified prompt, but a very, very good one. Models are tuned to follow your prompt, for good and for worse. If you prompt sloppily, you're going to get slop.

A spec sounds good. I don't know how often it will be followed in practice, because effectively what that transitions us to is a waterfall development approach, where you spend 3 days writing a 50-page document and kick off the agent. That doesn't seem right.

I obviously have some bias here, because Cognition has believed in interactive planning from the start. You kick off a thing, you get some feedback, and then you're like, “That's not what I meant. Let me correct myself,” because I don't know what I wanted when I started. You work with the machine to discover what you wanted, and the machine works with you to either get you what you wanted or show you the errors of your ways. Then you correct it from there.

One thing we talk about, which is very aligned with that thinking, is that there are 2 problems as these things get better. One is how you specify what you want, and the other is how you verify that what you got is what you were thinking.

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

It depends. If there are tests in your codebase, it's right out of the box here. Jules will run your test suite.

Exactly. But it's not like after everything there must be a matching test for the prompt that was mentioned. That would be the extreme of what we were talking about. I don't know if people always want that. Maybe it would be helpful to do that to show that it was right.

But let's say I don't write tests in my codebase. I want to merge a pull request that is introducing tests just for this one thing. I think in some ways the engineer should be able to control what kind of outputs they want. If it helps and they want it, absolutely.

Do you think there are other innovations in specifying apart from just chat?

Oh, totally. Agent IDEs, for example. Spec-driven development is in this category. One of them is multimodal. If I'm going to show a bug on our website, do I want to type out words to describe it, or am I going to point at the picture?

With Jules, you can upload images now. We have certain ways we communicate as humans that are easier in certain situations. Let's bring that to our engagement with the software.

Of all people, I expect you guys to be best at this because Gemini has video understanding. I just want to submit a video, because some things cannot be screenshots.

Yep.

It's more about the behavior of things appearing and disappearing. I would love that if you guys did it, because no one has it yet.

I know. I would love it too. I'll tag you now.

On my side, the vision—the version of that that we're exploring—is computer use.

Computer use was introduced by Anthropic, and then OpenAI dipped their toe in with Operator and now Agent Mode in Atlas. I don't know if you guys have done anything super splashy on computer use, but it's coming back. I can feel it. It ties into coding agents and just using AI systems in general.

Yeah, definitely.

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

Absolutely.

You need precision and speed, as well as an affordable cost.

Yep.

It's a lot.

Yeah. These are the kinds of projects that are so fun. There's just so much to build. I think one of the reasons you see so many companies in the space is partly that it's just so fun. There are so many things to build and so many tools that seem like science fiction.

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.

Yeah, it's awesome.

Okay, moving toward wrapping up: if people run into you at AI Engineer after they've heard your pitch on Jules, what else should they talk to you about? What can you help with, and what are you looking for?

Anyone should feel free to come up and talk to me at any point. I'm obviously very interested in anyone who's doing stuff with coding agents or using coding agents in an interesting way. I'm always curious about the workflows people have with their agents, whether it's, “I'm using this tool in this way, and I've configured this crazy thing.” I always love hearing how people are using it.

I also love hearing from people who are having a bad time with it. Maybe they're not coming to this conference, but I want to hear, “I've tried all these tools, I don't like them, I don't use them, and here's why.”

Yeah. I’m totally open for any side of it, all the way from full AI-pilled coding AI lovers to people who hate it. As far as what I’m looking for, I think I’m really just going to connect and meet people. We’re always hiring, so anyone who’s interested in working on this stuff, I’m always happy to talk. But, yeah, really just meeting people, spending time geeking out on this stuff.

Yeah, there’ll be lots of geeking out.

All right. Thanks for your time. Looking forward to it.

Yeah, same.