Cursor's Merrill Lutsky on Origin — rebuilding Git infrastructure from scratch for agentic coding at SpaceX scale

Speaker 2: Usually. Usually.

Speaker 1: Next, we have Merrill.

Speaker 2: This is a scoop. Getting a SpaceX employee on the show is difficult. Not many people can do it, but we can because we got Merrill. How you doing?

Speaker 1: Rocket man.

Speaker 2: Doing great. Rocket man. It's rocket man. Dweep, I I wanna go way back in time. Did you ever have the I wanna work for a rocket company as a kid? I wanna be an astronaut because you sort of wound up pretty close to it.

Speaker 4: Oh, yeah. I mean, it was I don't know. I remember watching like that. Like, do you ever watch that movie October Sky?

Speaker 2: Oh, yeah. I love that shit.

Speaker 4: Yeah. I remember. That was, like, that was, like, one of my favorite movies growing up.

Speaker 2: It's, like,

Speaker 4: a NASA phase. And yeah. So it's it's kind of crazy. Know? Somebody on Hacker News the other day commented, like, it's funny how adding trying to add StackDiffs to GitHub has resulted in the Graphite team now building rockets. Yeah. Here we are.

Speaker 2: Yeah. That is it's remarkable. It's like it's almost like a side quest in the in in in the long arc of history you'll be deploying on the moon and and Mars beyond. How has it been? What what what has the last couple weeks been like? Has it been sleepless? Are you able to still continue growing the product, building the product, or has all the news been infiltrating you? Like, what's the what's the psyche been like over the last couple weeks?

Speaker 4: Yeah. I mean, it's been it's honestly been really exciting on on the ground at Cursor.

Speaker 2: Like Mhmm.

Speaker 4: Everybody is everybody's fired up about about, like, the compute partnership and what that's unlocking for us on the model training side. We also just had our first big conference compile at Fort Mason yesterday. So the marketing team was was really working hard to get everything together for for that and make that a great event. And, yeah, it all kind of led up to yesterday being a crazy news day of the the SpaceX announcement and then, you know, the three big feature releases that compiled yesterday, including Origin, the Gentle Kit Forge that that we're working on. Yeah. So, So it's been a it's been a frenetic past few weeks, but everybody's really fired up.

Speaker 1: Did you guys expect the announced like the the sort of second part of the the announcement between you know, deal between Cursor and SpaceX to get as much like attention as it did? Because I was seeing people that seemingly hadn't seen the first the announcement of like the fur the first part of the deal. And a lot of people felt like they were processing it for the first time Sure. Which kinda makes sense Yeah. In hindsight. But in our little bubble, I was like, great. Like, obviously, they were gonna do this deal.

Speaker 2: Yeah. I sort of had the I guess I had it in my mind that you already have a SpaceX badge, but that was not the case because this was a partnership and now the deal has been announced. Is that correct?

Speaker 4: Interesting. Exactly. There was the the announcement of the option to acquire, and then the the other you know, the announcement yesterday was the exercising of that option. Yep. Now, I mean, I still don't have a SpaceX badge Sure. That, you know, where we have to wait for for the deal to close. Yep. Then we can actually start working together. But Yeah. Now now it's like just the just antitrust and all the all the other pieces.

Speaker 2: That makes sense. Well, congratulations. Do want to talk about See Walk me through the thesis there. I mean, there's been plenty of gripes about how much code is getting pushed to the Internet generally. How teams are working together, agents are working together. We're in a CPU crunch now. How much of this is just taking a developer workflow that already exists, speeding it up versus reimagining it from first principles? Like walk me through the thesis for this product.

Speaker 4: Yeah. I think the origin really started from a simple realization, which is that every single piece of the software development life cycle that we have today, every tool, every piece of infrastructure, was all designed for a world where humans write every line of code. Yeah. We've tried to scale that as best we can over the past year or so, but we're really seeing many of those pieces just crumbling under the pressure. And first and foremost, just having making from first principles, designing something that can handle the scale of agentic coding, that's yeah. That's the the most important thing is, like, is it scalable? Is it reliable? Is it performant? Like, can it handle hundreds of agents working in parallel at once? And that's really what we've what we've designed Origin to be is, like, the the first example of Git infrastructure that is really built for for the agentic era. The other piece of this is that then building kinda building on that, if we own the infrastructure, we can start to build it with agents in mind. So it's not just it doesn't have to be just like the git or objects. It can be you can start to store agent traces. You can see the whole history of every action the agent has taken, everything it referenced. You don't have to have it's kind of silly today that you'll use one agent to write the PR, but then as soon as the PR is opened, like, that agent just goes away. It's like, why does that not carry through and help you then, you know, address review comments, resolve CI failures, fix merge conflicts, and really, you know, we wanna get to this vision of of more, like, full self driving PRs that can really bring themselves out to production when in many cases without human intervention.

Speaker 2: Yeah. Yeah. That's interesting. Yeah. I I have a bunch of different things there. I guess the the question is yeah. How, like how important is speed in this context? How important is actually like the volume of code? You see a lot of these like someone will post like, okay, I just vibe coded this. I pushed this change. Whatever model they're using did really well. They got the job done. And then somebody will go in and look at the PR and be like 100,000 lines of code to change the color of the button? Like, this is evidence of slop or this is evidence of poor computer science happening. Do you have a stance on the is the answer to slop more slop? We going to get less sloppy, more terse in the writing? Is that something that's coming? Or do you need to design a system where the new normal is, yeah, changing the color button of color might be 10,000 lines of code change. That's fine. Let's also store the reasoning trace in the whole in the whole, you know, back and forth and everything that the agent was doing so that, you know, yeah, it was it was, you know, ten years of development time for one small change. It's fine because, you know, hard drives are cheap.

Speaker 4: Yeah. I I certainly don't think that the the future is 10,000 lines to change a button color. But I I do think that the volume yeah. Fundamentally, just the volume of code and the volume the throughput that we're seeing Mhmm. And all pieces of infrastructure have to handle now is is just growing exponentially. And Mhmm. You know, we've seen we we've seen now, you know, a lot of teams are are now doing, like, you know, thousands of of pushes per hour. Mhmm. You know, and previously, they were doing, you know, tens or hundreds. And that's a lot of what we've been been trying to make sure Origin is able to handle is is that scale. So we we've run some simulations recently where we're doing, like, you know, 80 clones and 22 pushes per second with no downtime. And, you know, it's it's obviously just this is just like us testing internally, but Yeah. It shows that, like, the architecture holds up and it's scalable and it's not yet the when your agents are ready to work at that speed, Origin is not going to be the bottleneck for them. But I I do think that we'll we'll then need to the other thing I think this will enable though is, like, more agentic review, more back and forth, like, more more autonomy and more more, like, spinning out of of changes and and refining of them before a human ever has to look at them. So, yeah, hopefully, we can avoid the case where you'd have, like, a 10,000 line PR that's that's doing a meaningless change.

Speaker 2: Yeah. How is how is growth overall? I feel I feel like a lot of people watch an acquisition happen or a series of acquisitions happen, and they're sort of like, okay, like that's the end of the story. Or they I mean, it's really easy with Elon Musk to look at it and be like, okay, well, like this is an incredibly talented team but this is data centers and space stuff. This is mass driver on the moon project. Like the main mission is now, you know, not relevant and so it's less focused. But I imagine that you're just seeing growth in the core business still because

Speaker 4: Oh, yeah.

Speaker 2: Teams are still, you know, getting up to speed on on pushing code and and using all these tools. What has growth been like post acquisition? What have been the big drivers? What what does that tell us about the way just the broad software development market is evolving?

Speaker 4: Yeah. I mean, we've seen since since the acquisition, we've seen, you know, continued growth in the in, like, the graphite business itself. But really, think the the most exciting thing for us has has been seeing the level of interest from companies of all sizes in in what we're doing with Origin. So Mhmm. We set a goal for end of July for, like, wait list sign ups that we wanted to hit, and we we hit that in the first twenty like, the last twenty four hours since we announced the product.

Speaker 7: So it's

Speaker 2: That Gong is also for the acquisition closing, which is maybe a bigger deal than waitlist. Honestly, like, traction on an actual product is probably more relevant than any financial milestone because it means that people are actually interested in this stuff, which is good. Yeah.

Speaker 1: It's Right.

Speaker 4: It's every time we've talked about it, you know, at at Compile yesterday, you know, on customer calls, it's like, you know, Compile got a huge round of applause. Like, it's just it it's so obvious this is a a problem that every single company is facing right now, and it's solving this this burning need. And that's why we're we're so excited about what we're doing. And, you know, really thinking to your point around around acquisitions, like, thinking back to to the Graphite days, like, one of the hardest lessons that we learned from building Graphite was that your product is only as good as the infrastructure that it's built on fundamentally. And it might seem obvious, but it's it's a really hard lesson to learn. And we are always limited by, like, wait and see, availability, you know, even just UI extensibility. Our customers felt those limits. You know, when when our infrastructure was down, we were down too. And the it's funny now. There's this whole wave of companies that are are going to their favorite agents and saying, you know, copy Graphite's poor request page. Make no mistakes. And, you know, shipping something out there, and they're really quickly finding out, like, why it's so hard to to do what we did. And even even then, I think we were never really able to to create, like, the the level of experience of just, you know, totally seamless to high performance interface that we wanted to because we were we were limited by the platform that we built on. With So Origin, you know, now with with in the backing of of Cursor and the combined team, we're finally you know, we're now able to, you know, truly realize and accelerate the vision that we had from the beginning of the Graphite days.

Speaker 2: Let's talk about AI doom, job losses. Is there any risk that AI puts DJing out of businesses? Are DJs going away in our AI future? Or will the DJ, the humble DJ, always have a place in our society? What do you think?

Speaker 4: I think there's always you know, it's always been about more than

Speaker 2: You've actually always been able to just press play on a premixed.

Speaker 1: You've always

Speaker 2: And yet we still are It's like, maybe

Speaker 1: you're ironically, like a really good White pill. White pill for just jobs overall. Yeah. It's like

Speaker 2: Yeah. You wanna see a human up there and and

Speaker 1: Even the most yeah. Even the most the most well known DJs, they have like a set that they could just go up there and play and

Speaker 2: dance around and of course of that. But anyway, great to catch up. Congratulations on the progress. Thank you so much for

Speaker 1: coming the for you. Happy for you.

Speaker 2: What an inspiring story. A true overnight success.

Speaker 1: That's right.

Speaker 2: 2020, baby.

Speaker 4: Overnight success of six years.

Speaker 1: Yeah. Yes.

Speaker 2: A lot of sleepless nights. But thank you so much for coming on the show. We'll talk to you soon.

Speaker 1: Great to see you, Merrill. Congrats.