← Return to Index Archived November 18, 2026
The Lead — Nov 18
THE PRAGMATIC ENGINEER · GERGELY OROSZ

Software factories and code reviews with Harjot Gill

CodeRabbit CEO Harjot Gill describes running fleets of AI agents that write, test and ship software at a pace that has made human code review obsolete. He argues that engineers must instead manage risk, production quality and the agents themselves as token and computing costs reshape startup competition.

1h 26m / November 18, 2026 /aistartuptechnology / Transcript sourced from openai
All episodes from The Pragmatic Engineer →·Episode page →·Podcast website →·Listen on Apple Podcasts →

Overview

Harjot Gill, CEO of AI code review startup CodeRabbit, talks with the host of The Pragmatic Engineer about running his company almost entirely through AI agents. Gill says he merged more than 5,000 pull requests (PRs) in a single month without opening a code editor. The conversation covers what he calls "software factories," why he thinks human code review is finished, and his claim that a large seed round no longer covers the cost of building a software startup.

Key Takeaways

Gill splits AI work into two modes. In one, he steers an agent message by message, like shaping clay on a potter's wheel, which he reserves for UI work. In the other, he hands an agent a goal and lets it iterate against tests, simulations or production signals until it gets there. He says about 18 of his 20 active sessions are the second kind. One ran for six weeks and produced a Rust storage system now in production, and he says he hasn't read a line of it.

His view on review: humans have stopped reading diffs, but someone still has to judge risk. At CodeRabbit, the unit of human review is now a release, which ships every 30 to 45 minutes. A Slack summary rates each change's risk, and engineers decide whether to roll back. Accountability stays with software engineers and site reliability engineers (SREs), not designers or product managers.

Engineers hold a "driving license." CodeRabbit traces production regressions back to the PR that caused them. Repeat offenders lose the right to ship without review and have to earn it back.

Gill's approach to quality is to let entropy rise and then run cleanup agents. These remove flaky tests, prune redundant unit tests, and refactor bad abstractions. He also uses formal methods, the TLA+ specification language and the Lean theorem prover, to catch race conditions before code is written. He claims defect rates have gone down as output went up, and argues that humans wrote more slop than validated AI does.

Costs moved. Gill estimates his personal token spend at around $250K to $275K a month. He says a vendor quoted roughly $13 million a year for continuous integration (CI), so they built their own. They are now running into CPU shortages on Google Cloud.

On funding, Gill argues that copycats now appear within weeks, so venture-backed startups need more capital to buy tokens and compute rather than hire. The host pushed back hard. In his view this holds mainly for AI dev tools companies in crowded markets, not startups in general.

Practical Steps

  • Give agents ways to check their own work close to where they're coding: local deterministic simulations, API emulators, then real cloud dev environments.
  • Roll out behind feature flags in stages and have an agent watch logs and traces before widening the rollout.
  • Assign agents ongoing ownership of specific systems, such as billing health or flaky tests, and have them add telemetry or open fix PRs after each deploy.
  • Track regression exposure time per PR and tie review requirements to each engineer's track record.
  • Measure cost per PR to spot token spend that isn't producing work.
  • Move long-running agent sessions to cloud sandboxes so other people can unblock them. Gill currently wakes at 3 a.m. to do this himself, which the host said he would not recommend.
  • When hiring, look for side projects, breadth across domains, and comfort juggling many parallel tasks. Gill half-jokingly screens for hundreds of hours of strategy games like Civilization or Factorio.

Notable Quotes

"The human code review is dead, but the human review is not dead." (Harjot Gill)

"Humans were writing slop, not AI... AI is actually writing less slop at this point, if it can validate against something." (Harjot Gill)

"Burnout never happens due to hard work. It kicks in when your work does not lead to the desired outcome for a long time." (Harjot Gill, quoted by the host from his writing)

Burnout never happens due to hard work; it kicks in when your work does not lead to the desired outcome for a long time. — From the episode

Full Transcript

Source: openai 1h 26m runtime

What does it look like when a startup CEO merges 5,000 pull requests in a single month thanks to going all in on AI agents and software factories? This CEO is Hardjot Gill, heading up CodeRabbit, the AI code review startup. Hardjot had been writing code for 30 years, is on his third startup, and never shipped as much as he does now. Today we cover why Hardjot says human code review is dead, but AI code reviews are more important than ever. The concept of driving license for engineers, how devs at his company lose their right to ship to prod if they ship too many bugs. His spicy take that $8 million seed rounds are no longer enough to build a software startup, and me pushing back on it, and many more. If you want an unfiltered look on how a truly AI-pilled founder slash software engineer operates, who is extremely hands-on with AI, this episode is for you. And while I am nowhere near as AI-pilled as Hardjot, I found it educational to understand his thinking and his views on the dev tools industry, where competition feels more intense than ever before. This episode is presented by TurboPuffer, and I want to show you the new and improved the Pragmatic Engineer search at pragmaticengineer.com/search that I built on top of it. I was frustrated with Substack search, how their search was keyword only, and I couldn't do a search like the article where me and Ken Beck made fun of a big consultancy. I'm obviously looking for the piece about McKinsey, but Substack cannot find it for the life of it. So I decided to build a search that I needed. I did the whole thing from idea to deployment in three hours with Codex and TurboPuffer, with Teepof doing the heavy lifting. I dumped all my articles into TurboPuffer. It handles creating embeddings and getting full text search working. The result is a much better search than what Substack has. Check it out for yourself at pragmaticengineer.com/search. But let's think bigger. What would it take for Substack to give every single publication this kind of advanced search? With most other search systems, you would need to create a bunch of shards, and it would be expensive to operate the search indexes. But with TurboPuffer, each publication is simply a namespace. Idle namespaces live on object storage, which has no compute cost. Active ones are pulled into NVMe and memory caching tiers, so they are served very fast. And this is how TurboPuffer can scale to hundreds of millions of namespaces, and why multi-tenant AI products at companies like Ramp, Harvey, and Notion are built on top of it. With TurboPuffer, you really do get the best of all worlds: ridiculous scale, low latency, and cheap operating costs. Whether you're building a side project like I did or a platform with millions of accounts, check it out at turbopuffer.com/pragmatic. Harshad, welcome to the podcast. Hey, thanks for having me here. It's great to have you here. This is your pull requests per month. Like, the last month, it was more than 5,000 pull requests. That's right, that's right. How? How? At this point, like, it's like really machines were building machines. Like, I mean, just good at keeping the agents busy at this time, I would say. What are these pull requests doing? When are you building it? Because, you know, like, you will have a lot of responsibility and time spent just running the company, I guess, coordinating people, pulled into fires, the usual stuff, right? Yeah, I mean, it all started, like, I think when the models, I would say, GPT-3 Codex was kind of the first inflection when we started seeing these agents getting good enough to do, like, real-world work. We can even ship to production. And around that time, like, it was very clear to me that, okay, we have been around for, like, two years. Like, I mean, so generative AI has been going on for a while. But last three or four months were, like, a big inflection point. I also see it here. It was up to July, you had, like, I don't know, maybe 100 or 200 max, and then it went to more than 1,000 in July. It was just something else, and you wouldn't believe, like, everything just breaks when you start shipping at that scale. So as a leader, like, one of the key things that I keep doing is, like, how do I make my team productive? And a lot of my pull requests go in just fixing some bottleneck in our SDLC at this time. I mean, of course, I'm doing, like, maybe one or two tactical features where I'm driving the agents with each message to craft the UI or something. But most of my volume of the code will be, like, believe it or not, fixing things like janitorial work, like fixing CI or flaky tests or building some, like, distributed system, like fixing something in caching or something, right? So all of that hill climbing is what I'm doing with these pull requests. Right now it's not very optimal. I would say that numbers should probably, like, be 10x of where it is. We're not even hitting the peak here with what these agents can do. And what agent or what tool really kicked this off around, I see, I was July, August? Was this Opus? Was this Codex? Was this something else? Yeah, I think, yeah, both the model providers have been, like, in lockstep in terms of capabilities. So a lot of our team members use Opus 5.5 with Claude Code, but I personally use Codex. I mean, I've been a big fan of the Codex desktop app. Like, I've been, like, coding, like, 30 years. I've done three startups, and most of the time I've also been, like, kind of the CTO hat in every company I've been. And I have been in NeoVim person. I've been a code editor person all my life. Even, like, had put out my dotfiles, which a lot of people use. But now we are in a different zone altogether. I have not been in a code editor since I shifted to Codex desktop app. It's pretty awesome. And ever since then, like, it's been always about, like, how many agents can I keep busy at a time. So I'm actually not sleeping much. That's the other side of it. Like, you get all this anxiety that you have to really— because the kind of value you are creating, I mean, if you know really what to build, then the kind of value you are creating is phenomenal. Like, I mean, I've never seen things move in such a compressed time frame quickly within the companies. So it's not just the volume. Like, these are not, like, some silly PRs. They're, like, actually moving the needle forward for our company. Okay, you're not sleeping much, so it's eating into your nights, clearly. And you're running agents there mostly with Codex and some other tools. But when you're commuting, you know, a lot of people listen to podcasts or music or just drive, and you have a cheat code there, right? Yeah. So I kind of live in the East Bay, and I take the Bay Bridge on my way to work here. And I have a Tesla FSD, and I love FSD. I'm, like, roughly 97% self-driving on FSD. Like, pretty much it's like driving on its own. And what I do is, like—and FSD has gotten really good. Like, earlier it used to, like, nudge you that you have to hold the steering wheel or something. But now you can, like, as long as you're, like, looking at the road every few minutes, it's okay. So as a cheat code, what I'm doing on this side is, like, I have my laptop with the Codex session, and I'm just, like, sending the instructions while I'm driving. So it's kind of funny, like, the AI is driving the car while, like, another AI is driving the company. Wow. Which I'm stealing. It's wild that we're there, right? And also, I assume, you know, you can now talk as well, or like, you know, there's Simon Willison was telling me how I think a year or two ago he was already talking. It's just crazy how different hands-off or different areas you can just use AI to, like, build stuff. Yeah, that's right. Like, one of the things with a company like CodeRabbit is, like, as you start shipping faster, every part of your business also have to move at agentic speed. Like, what we are finding is we're shipping so much that the attention, the marketing has become a bottleneck. So I had to go back and also, like, refactor the marketing foundation to be very agentic, which means that we built, like, a separate repository. We call it the content studio. All our slide decks are now in code, HTML, CSS. We built a whole portal to, like, share those, download PDFs, and so on. We actually are running even ads through code. Like, we have millions of dollars of Google ad spend, and the way we are going after all these keywords and the different copies on that is we have a loop where we look at the ads performance, and then some agent comes up with the code changes, and we actually create pull requests to deploy ads. So when you merge that pull request, it deploys it on Google Ads. So it's like everything is code, inside CodeRabbit at this point. Okay, so you've been in the industry for close to 30 years. I just want to go back before we get more into, like, how we're building now, but how did you get started? How did you get into the industry? Where did you first work? Yeah, I mean, I started my career in very early stage in a telecom company, which was actually a startup acquired by Alcatel-Lucent back in the day. That's where I first got a way for people could start companies and build deep tech. I mean, that was like carrier-grade systems, high-availability systems. So I learned a lot about distributed systems in my first job. At that point, I kind of like felt that, okay, I want to do my own startup, and that was the reason I came to the U.S. And the route I took was, I always wanted to do a startup. I mean, that's very clear. It's not like something— Very clear. And now you're on your third one anyway. So the route I took is I came into my PhD at University of Pennsylvania in the East Coast at Philadelphia. And the goal over there was very clear that, I mean, I'll incubate some good idea that I'll take out and then commercialize. And that's exactly what I did. So we did a technology transfer from the university. My first startup was kind of also like an inflection point, right, when like microservices and Kubernetes were taking off. We saw that the observability is going to become a bottleneck. Yeah. So that was my first startup. It was called Netil, listen spelled backwards. Oh yeah. And this was observability for Kubernetes, right? Monitoring. Microservices. So the idea there was like these, prior to microservices, you had monoliths, and you had like a Java monolith where people will install an APM agent like New Relic, AppDynamics to instrument that. So your function calls were in the code itself. And when function calls became API calls when you had microservices, so you started decoupling. Then you had to like really observe the network. Your network became the app in a way. Yeah. So that was the idea behind Netil was, can we listen in to these network events to understand application performance? And that was kind of precursor to eBPF and some of those technologies that came out later. After the startup, you moved to a larger startup, right? Or did you move, or was it an acquisition? We got acquired. It was a— By Nutanix, right? Yeah, this was a public company at that time, and we got a pretty decent offer. Like we just did one round of funding, but then the offer was pretty good that we took an off-ramp. Then I stayed there for like a few years trying to like spend time integrating our technology into their stack and the product offerings. Then I quit to start another company called FluxNinja. Which was in, like, web scale, like how do you do load shedding, graceful degradation, controllability, right? So we built, like, whole PID controllers to make sure that the applications don't crash with cascading failures during, like, Black Friday spikes and so on. Now, very interestingly, like, at that time the market had appetite for all these Kubernetes service meshes, but then generative AI happened all of a sudden. We saw that first, yeah, when I touched GitHub Copilot. I mean, we had TabNine. This was 2021, right? Yeah. So at FlexNinja, we were using some AI, like TabNine is one of the products we were using. Oh, TabNine was the OG of AI autocomplete. Yeah, and it would, like, create its own local model or something. I don't know what it used to do, but it used to work, like, but it wasn't, like, as great as, like, when the GitHub Copilot, that was, like, the aha moment, right? Yeah. And this was, I believe, even before ChatGPT had come out. At that point, it was very clear, like, we were a fast-moving team, and at that time the code review was already a bottleneck. So people were doing stack pull requests. We built our own tool called SPR. We open sourced it. But the thing is that, like, it was still not ideal. Like, given the volume of the code just started growing after GitHub Copilot, it was very clear that AI is going to write most of the code. And so just to pause there, when you saw GitHub Copilot in 2021, this was before ChatGPT, you already predicted or kind of, like, fast-forwarded, saying, Oh, AI will write most of the code? That's right. This was not a commonly held view. Like, a select few might have held it, but what made you assume this? Can you talk us through your thinking process at that time? I was in the arena, like, as a leader. Like, I was actually coding on the ground. Even at that company, I mean, a lot of the work I was doing was janitorial, like just making the developers more productive, adding all the linters, the tooling, everything to go faster. Right, as a leader, like my goal has always been like how quickly we can get to the next milestone, right? And when I touched that system, I don't know, like it just came to me that, okay, this is going to change everything. The tool was really good in like sensing what I needed to do next, and I was accepting the suggestions. Like it was like writing really complex algorithms. You just provide a few comments on what you want to do, and the code came out, tab complete, right? It was pretty magical. And at that time, it was just clear to me, I don't know how, but I mean, once you are in the arena, I think you see these patterns long enough, the second-order effects. And the same thing was clear to me three months back on how software is going to be built. Like I would say I'm still learning. It's not like we're not on a stable ground yet. Things are moving too fast, and some of the things I'm seeing now, we could not have predicted even six months back on how the software development is going to evolve from here. You found a code rabbit on this kind of notion or prediction. Was that how it happened? Actually, we didn't realize this would be such a big opportunity. One thing was clear: it will become a bottleneck. But we thought, I mean, the market cap, we were looking at Snick and a few other companies in the GitHub ecosystem, right? Because we have an installation which is like inside the version control systems like GitHub. Like there was some precedents around previous generation companies like SonarSource, based out of Europe, which are like code quality agents. We were kind of like thinking, okay, the market cap would be like in that territory. What we didn't realize is that this is like a much, much bigger market cap opportunity, because people who were not even traditionally buying those tools would line up to buy AI code reviews. Even though we don't sell to individual users, a large number of our users are individuals paying for the product. So you've been a NeoVim user. You wrote a lot of code for like decades, pretty much, right? You're handwritten. How easy did you find it to just move over to AI? Because a lot of people, especially some folks who use NeoVim, you know, like they're kind of into the craft as well, they might have been a bit more hesitant to just go there because you kind of have to, you know, get rid of that really familiar UI. You no longer even have a UI when you're using agents. So for me, like writing the code is just means to an end. As an entrepreneur, like you want to get to the final value and validation quickly. I mean, I did enjoy writing code, but I wouldn't say that was the most enjoyable thing in building a company. It's about like how quickly you can get an idea out, get validation, grow that business, and so on. Coding was kind of a necessary thing you had to do, and a lot of the work around maintaining those codebases, DevOps, all comes alongside that, right? So for going from—I don't miss it. If you ask me, like, do I miss writing code by hand? I actually don't. What I love is now I'm able to go from ideas to very quick implementation and just validate it. So I'm doing actually the fun part of software development. I'm still doing software development. It's not like I'm not doing software engineering. But now what I'm focusing on is a lot more fun than figuring out the syntax and autocomplete. I mean, my agents are— I mean, my agents are writing Rust code. I have zero idea how Rust works. We recently, like, had an agent goal running for six weeks to— Six weeks? Yeah. So for running, so building, like, get write head log on object stores. Like, when cursors origin blog post came out, I pointed the agent. It ran for six weeks, and now we have that in production, powering a lot of the knowledge bases at CodeRabbit. So I didn't even have to have the real implementation. It just hill climbed over time on its own, all written in Rust, and I have not read a single line of it. And for those of us who are not as familiar with this specific term, what does hill climbing mean for agents? Yeah, so it's some sort of a loop. Like, it all starts at a very basic that agent can verify its work against something. Yep. And these loops are like inner loops, outer loops, multiple types. Like, so when you're looking at software development, you want to, like, provide agents means to verify its work close to where it's doing the software development. So we do a lot of, believe it or not, a lot of formal methods. So when we first did the Slackbot product, like eight months back, we invested a lot in, like, first modeling what this distributed system is going to do in TLA+, then implementing it, and also investing a lot in simulations. So the way we are scaling Code development—Code Review is just one piece which we, of course, market, but there were like a hundred other things you have to do, right? Yeah. So we invested a lot in local deterministic simulations, right? Very recently, we started investing in Lean theorems as well. So a lot of that is like— Lean is another form of formal methods, right? Yeah, like you have theorems, and even, like, used for mathematical proofs, but you can also try to build provably correct software, things like billing service and all. We have got, like, a bunch of Lean theorems. I can actually show you after the podcast. It's pretty cool what we are doing there. So what we have been doing is like— These are all like hill climbing. Like you're first making them hill climb locally, then you're going to the next environment where, let's say, emulators is another thing we have invested a lot in our own. So you can now very quickly mock a real emulator, like the real APIs, like Google Cloud APIs or whatever, right? Then you are actually testing against a real Google Cloud or Vercel environment. Those are the cloud providers we have in dev. You can have the agent actually spin up services, actually test them, right? I mean, this is all hill climbing. Then in production as well, like the thing is that I kind of see each release. So we release multiple times a day, right? Every 30 to 45 minutes, we ship code. And each release is almost like a turn in a strategy game. If you have played Civilization, you end your turn, you learn. That's how we see releases. Like each release is an opportunity for an agent to learn something from that environment. So a lot of the times we will release features under a feature flag, do like progressive rollouts, like maybe 10% for our open source users and so on. And the agent will observe the performance, the logs until it's stable, then do wider and wider rollout. And each release is the opportunity to learn for the agent. Interesting, because what you're describing, this kind of stage and wider and wider rollout, this is something that Facebook had been doing for a long time, and they just have a massive user base. We're talking billions of people, right? And they built this automated system. And outside of Facebook, I don't really know, I didn't know of many other companies that had a kind of automated rollout for a few reasons. One is this only works when you literally have hundreds of millions of users. It's really hard to do it otherwise. It's a large investment to build. Automation is hard to build. You need, even pre-AI, you needed some sort of ML, which again just goes up and up. And now you're describing that you, at a pretty small company in terms of headcount compared to Facebook's 75,000 people, CodeRabbit is like 40 engineers maybe? Yeah, and you've been able to build this for yourself. That's right. So the way we are seeing it is like the way Netflix, Facebook used to ship with thousands of engineers and the kind of tooling they needed, like all the way from... The process around merge queues and everything, like they needed to ship at scale, is something that small teams are going to need now with AI. Now, teams of 30, 40 people, if they use the agents in the right way, have the right abstraction and this whole concept of software factory, they're going to ship as much as Facebook, which means that all the sophistication they had in terms of testing in production, managing the risk, managing your CI/CD, everything, caching, there's a lot of stuff around it. I mean, they would need it in a small team setting as well. So it's a golden opportunity if someone was at Facebook at that time, like when they were shipping and figuring out all these concepts, to go out and now actually build products. The market's grown. There's a lot more people who need those products than ever before. And you mentioned software factory. It's a term that is going around a lot. What does it mean for you? Yeah, it's many things. I think, I mean, we have a slightly, maybe it's the same take as others, but I mean, I'll give you my take on it. Even when you're building with AI, there, I think, two ways of using AI. One is more like on a potter's wheel. You are shaping the software with each interactive chat with Cloud Code or Codex. Like, for example, with UI development, I have the browser open, I'm doing annotation, and each time I send a message, I'm crafting the software, but with AI. And I can manage as a human, my decision capacity is to do only one or two of those tactical chats in a day. I mean, I usually have like 20 chats in Codex right now. Two would be of that manner. Around 18 are more like the factory concept, like where I know that I can have the agent hill climb on something, whether it's like building a distributed system, I can be hands-off with a goal, just like the example I gave you with the Git on object stores and also SQLite LTX. That's the other thing we did with knowledge bases. We removed our vector database provider to build our own in-house version on SQLite LTX, right, on an object store. Those kind of things you can hill climb. The other stuff is like each time you have a deploy, CI/CD is happening. It's an opportunity for the AI to learn and continuously improve it. So I have agents running continuously to remove flaky tests. Right, each flaky test resets the merge queue. It's really bad. So you have to have that continuously happen. Each time you have a deploy or a CI, each run is an opportunity for the agent to optimize something, do the next thing, right? And then you're running the agents against production as well. So you have agents which own health of some components I care about. For example, I care about billing. I care about the knowledge base performance. Like I care about the code graphs we have. So each deploy, I have the agent running in a loop to observe the next experiment and do either two things, like add more telemetry if it's like flying blind or fix something. So those are the kind of loops I'm running, and those are factories. So the thing is that right now the chat is not kind of the right interface to run those agent fleets at scale. The chat is great when you are doing crafting the software one message at a time. So some of these things we are still figuring out. First of all, that whole maintenance, the janitorial work around shipping software and maintaining the factory should probably be in the cloud on sandboxes and all. And I shouldn't be the one waking up at like 3 a.m. I mean, it's more like morning shift, night shift now in the factory. We luckily have a team in Bangalore, so we have a follow-the-sun model. So I want, like, if I'm babysitting the agent, so I'm not owning the code anymore. We are owning agents. Let's say I have an agent looking at the health of a billing service and it stalls on something, it has a question. Then if I'm around during the U.S. time zone, I'm able to answer that question. But I want someone to be able to unblock the agent while I'm sleeping. So I would say we haven't figured it out yet. So we are going on the upward trajectory, but most of our development right now is constrained by how many sessions people can run locally, how many releases they can, like, make their PRs land in. Like, look at the release coverage and the PRs they landed. And we're also constrained on local resources. One of the things we found, and that's where I have a cheat code as a CEO, I have a very powerful machine. The developers don't have those. Like, they have like 48 GB RAM, 24 GB RAMs, and they max out at like three, four sessions. Beyond that, the machine just dies with the unit test. It has to run locally, right? So we are feeling like the real compute constraints right now in terms of the local development. And we are investing very heavily to go and make some of our cloud-based environments work. But I guess it's pretty obvious at this point that a lot of development will move to the cloud because of a lot of the constraints that you mentioned. And I guess the only downside I can think of moving to the cloud of running CodeRabbit in the cloud is, A, there's a bunch of work that needs to be done that you have an environment that's similar to your local one, security, all that stuff, but you need to build that. And there's potentially a cost question as well because now, you know, like you have your hardware, and when you're utilizing it pretty well, that's pretty good, right? You buy it once and you do it, and there will be an additional cost part. But the cost, I wonder if that's a smaller question because, you know, like we already know AI has very real costs, and I know you're paying a lot for AI already. It's actually better because in the way, like if it's in the cloud and more like a collaborative setting, you clearly see the value created because right now we have some people who are spending tens of thousands of dollars on local tokens. We don't know if they're doing like side projects, like everyone's building their home automation systems. I see what you mean, yeah. Right? So there's no auditability as well on local development, right? So sometimes we look at the price per PR, like mine's, I don't know, like 150 bucks a PR. Like for some people it might be thousands. Then you have to question, like, are they really doing your work or someone else's work? So the thing is that, like, the cost perspective, I think I will be also, as a leader, comfortable spending the money if I can see where it's adding the value. Like if it's in the Slack, so we have a Slack bot over there when CodeRabbit responds to all the Datadog alerts, I clearly see it's adding value. CodeRabbit, one of the main things that you do is AI code review. Now I have to ask you this question: Is code review dead? Because specifically human code review, there are so many signs that it is on its last legs. The teams who are still doing it, who I talk to, it's kind of theater. It's like, you know, LGTM looks good to me, but people no longer add comments. They no longer read it. And companies like Uber are now making changes to just make it so much easier for people to, you know, like see what a code review does, because they also know people no longer, they spend less and less time reading it because there's just more of them, right? You're right. I mean, the human code review is dead, but the human review is not dead. Some review still has to happen. If you look at the software development space, there used to be like an upstream kind of a guard, like when you were planning the work, because your engineering was an expensive resource. So when you were using the tools like Linear, Jira, one of the reasons to do this agile and sprints was to make sure the developers are working on most important tasks. We don't, like, spend a bunch of time and effort building something that will throw away. Exactly. So there was a lot of regulation upstream earlier. Now that floodgate has suddenly opened with AI. Now you can go from idea to code without going through planning. Like you have a support ticket coming in, people will just ask Devin to look at it or codeRabbit, Slackbot or something, right? At the same time, your designers and people who were not software engineers are now doing coding as they should be because they know what to build. A lot of designers are shipping code. You have background agents like Datadog is now going to be opening PRs. Each time it sees an alert, it's going to say, I'm going to do it for you, right? Yeah. So what that means is that your planning layer and the review gate still has to exist somewhere, and that becomes a pull request because everything upstream, there's an influx now happening, right? And when you have a pull request open, it doesn't mean that you have to ship it. At that point, you decide whether this is worth spending my time, whether I take it to production or not, because the human taste is the bottleneck now, human judgment. So when a human is coming in, they're first making a decision: is this something I need to ship? Is it slop I need to close? I mean, and you make a decision after the code is there, right? Second thing you're doing is like, what's the risk? Can it just go through production without me looking at the code, or do I need to understand the broader implication? Is it like a database migration it is doing? Why is it doing the migration? Because you want to sometimes look at the schemas, because that's more structural change, right? So we do all that, like risk management within CodeRabbit. So some PRs are on fast track to production. We are even letting our interns ship directly to production. So human code review is dead, but I would say human still review some of the higher-order abstraction. And how we are, like, doing this with SOC 2 compliance is also interesting. So we don't really— so the definition of change at CodeRabbit is now a release. So we have a Slack message. We kind of get all the indicators around the risk we are taking. Each PR, you have a minimal risk, high risk, and over there you kind of make a decision on whether this can go or some things have to be rolled back. So each release is what we are reviewing now, not PRs, as a human. And then when you say human review, just to be clear, human means software engineer, or can it also mean non-software engineer? I'm just curious if you still think the accountability needs to sit with people who kind of know the internals, or can it be someone else? Can it be product manager, designer? I think it's still software engineers at CodeRabbit. Like, they're not necessarily looking at the code anymore, but they are actually understanding the risk that we are taking with each release. So it's about who can understand the risk. Yeah, because if something goes bad, then they are going to be the one dealing with it, like rolling back. Like it takes like 10, 15 minutes to roll back. I mean, that's the worst that can happen if something goes bad. But they are still taking the responsibility. So we ideally want, like, developers to be around when something goes out so they can babysit what they just shipped. Yeah, that's why we try to release so frequently. But at the same time, like, with the person who's approving the release, we have SREs, and also they also, like, are taking some responsibility at that point. I think it makes perfect sense, especially for a business, because we're talking about a business where you actually have revenue, you have users. You know, like if slop goes out and it takes down production, you're gonna just lose customers or goodwill or cost frustration and so on. So I think it's great that you're saying, okay, let's have someone accountable. But the problem that us devs have is there's just a lot more changes and— I'm not going to have time to read through all of them. What ways are you thinking, either now or in the future, what do you think needs to change so that as engineers, as the person who's making the decision, I can just have better information to, you know, like focus my attention where I need to focus? Like, for example, I don't really want to review a manual refactoring that just move things around and it compiles. I mean, I could, but, you know, there's good use of my time and there's poor use of my time. And I feel, as you will know from yourself as well, I mean, you've shipped like so many pull requests, you will probably focus your time on the critical stuff. That's right. I think the human judgment is scarce now, right? And you want to, like, first of all, not look at the code. There's higher-order abstraction. You want to understand the risk, understand the blast radius, especially if you are familiar with some area of the code. You might also want to have inputs on the architectural change. Like, for example, like if you're rewriting a migration on a large table, like it's going to crash in production. You have some of these, like, tribal knowledge you want to bring in, right? The other thing is, like, also, like, I kind of see, like, now driving these agents almost like getting a driver's license, right? So we basically even on the developers put some responsibility on them worshipping. So we let— So we let people ship if they are in the, like, review exempt list. So we also have, like, different lists and groups and teams where, based on their experience and their past performance, they can either go to production without a human review, or we put them in a group where they need a human review. That sounds like based on a trust that you built up with them. The way we have done that is, like, we do look at how long there was a regression in production. So because we look at the code, when you fix something, we are able to tie that back into the PR that introduced it. And now if someone has been irresponsible, like they did hard prompting there, or they did not follow through their change into production, you will have sometimes a hundred hours of exposure. Means someone had to point out, like a customer had to point out that something regressed. You didn't QA it or something. And once you have, like, enough strikes, we take your driving license away. We make you go through the code review. Then you go through, like, a proving period with some human reviews, and then we let you ship on your own. So we're trying to, like, set up a culture where, like, I think responsibility is also with the developers to not drive these agents off the cliff. I like this driving license analogy, honestly, because we talk a lot about—well, I talk a lot about or worry about quality, because I can see that quality seems to have dipped across the industry, and clearly it's because there's a lot more things going on and a lot less review. And it's nice to tie back, you know, that feedback loop, that look. Like, I think it's human, right? Like, the more you ship, the more likely you'll mess up. And as long as you own it or there's consequences or you say, like, okay, from now on, for this time period, either, you know, come up with better practices or have a second person, have a second pair of eyes, which is expensive, but this is what we used to do even before AI, right? The intern, you give the trust to ship, but we didn't have that. The intern needed to have, like, two sign-offs, for example. Right. I mean, it's almost like that. Like, I think people have to have responsibility. Like, we have seen some developers are really good in asking the AI the right questions around the implementation it just generated. They're more diligent. And some developers will come and give, like, an ambiguous high-level goal. They will say, just solve this problem for me, and AI will, like, completely re-architect it. Sometimes you have to pull the plug when we start shipping, seeing those PRs. Oh, we are saying, okay, we need to stop that right away. We have seen those kind of things. So some people are, like, good; they've been prompting and steering these agents, and others are just reckless. And that's a behavior we then try to penalize. Yeah, I do like the idea of responsibility as well, but there's a question of, like, why, if the AI will get even better, right? Like, we're just speculating here. But at that point, why would you pay software engineers? And I kind of keep coming back to the idea that, well, you know, why do I pay a plumber to come when I could just watch a YouTube video and I could fix my toilet? The problem is that, you know, like, nine out of ten times it probably works, but the ten times I might either clog it or break it or something, and I'm paying the professional because if they come in and they mess it up, they are responsible and they will fix it. And I'm kind of thinking that as software professionals, you know, like, we will be responsible. And, you know, one of the reasons we will hopefully be paid the big bucks is we can take responsibility for a complex system where an outage is just very costly, right? No, I think the responsibility to some extent matters. Like, the things you can't foresee, we all have to be reasonable with the responsibility. Software are still, like, complex systems. They break in very complex ways, right? Fair. And you saw that today, the tweet around, like, railway employee hacking and then blaming agents going off track. But in the end, like, it's really humans who have to be responsible with how they're driving these agents, right? And they can do your bidding, but they can also go in crazy directions, right? I think that's important. And the other thing is, like, with this kind of, when you're shipping at agentic speed... Your operations also have to be agentic. So one of the things we are, like, agreeing is that, okay, our goal is not to prevent every outage. It's like to make sure that the exposure time of something bad happening in production is as little as possible, which also means that agents need to be looking at production. Yes. If that's out of lockstep, you can't ship fast. You're going to be more like AWS. You burn your hands with AI, and you're going to say, ban AI from my codebase. What you're saying, as I understand, is, okay, you know, we'll have a Jentic feedback to verify the code or give AI code review, whatever that is, or run maybe fuzz tests, et cetera. But the observability layer should also likely be tied. Like, it's all part of the QA. It has to, because your defect rates maybe are not going down, but now you're shipping more defects because the volume has gone up, right? Which means that you have to stay on top of them. Like, you have to have agents always looking at not just alerts. In fact, alerts are limited by your imagination on the failure scenarios you imagine. Like, actually, these agents are really good at looking at logs. They're great at looking at traces, right? So the way, in fact, those are the areas we are still not, like, fully scaled yet. Like, there's so many areas of our code where we have had longer exposure in production with bad code because no agent was looking at it. Wherever we have agents looking, we are able to quickly resolve stuff at agentic speeds as well. So entire software development from left to right is now going to be driven by agents, including production, and including marketing. Even stuff going out, you have to now be agentic. Hardik just said how software development is going to be fully driven by agents. And when agents write, ship, and fix most of the code, you want to know why each change happened, which brings us to our season sponsor, Entyre. A few weeks ago, I told you that Entyre was building a Git forge. Well, it's finally here. You can now host private repos directly in Entyre in the US, Europe, UK, India, Australia, Brazil, and Singapore. And Entyre has just reinvented the pull request for the agentic era. The problem with PRs is how agents write faster than how any of us can review it. So Entyre replaces it with something called trails. A trail connects everything about the work in one place: the intent, the code changes, the agent sessions. A trail can be spread across multiple repos, so the changes to your backend and your web app gets reviewed once. With a trail, you can configure custom evals. They look like heads-up display on an airplane and score your work on every push. And you can see how each agent run passes or fails your user-defined gates and decide when you merge based on these. The point is that you control when you allow agents to be allowed to merge the code, like only merging after a human review or after all gates pass. Oh, and one more thing: entire ship continuous memory. Entire turns your repo into a brain so your agent can search past coding sessions to find the context behind earlier changes, basically answer the why behind code changes. Memories get shared across the team and across the repo. If you want to see what an AI-native Git forge looks like, head to entire.io/pragmatic and check out Trails. And once agents are observing production and pushing fixes, someone has to decide what they're allowed to touch, which is why I'd also like to mention our season sponsor, WorkOS. WorkOS provides the auth infrastructure B2B and AI-native companies need to sell to enterprise customers with SSO, SCIM, RBAC, and now authorization for AI agents. WorkOS Airlock is intent-based access control. Most access control only asks who you are. Airlock also looks at what the agent is trying to do. At the moment, it tries to do it. Clear-cut rules like no agent deletes cloud resources are run as plain code. For the fuzzier calls, an AI judge evaluates the policy at runtime. You write these policies in plain English, and Claude can pull the ones you already had out of Notion or Jira and import them for you. Your agents, whether that's Claude or Devin, connect to a single access point, and your MCP servers are behind this access point. Airlock is a new layer on top of your existing permissions, which means that an agent can never get more access than the person who invoked the agent has. When something risky comes up, Airlock can escalate to human for approval, say over Slack. The idea is to write the rules once and spend human attention only on the calls that need it. You can request early access at workos.com/airlock. And with this, let's get back to Harjot and what happens to code quality when agents write most of the code. Let's talk a little bit about quality, because in your previous startups, you know, pre-AI, you built some infrastructure tooling, which, you know, reliability is very important at those places, accuracy when you have the monitoring. I assume these were all non-negotiable. But what's happening now with AI, what we're seeing is when you're using AI to generate code, it's great. It does it faster than I could do, usually similar or better quality than what I could have done. But it does get verbose. It can just repeat itself. It can just add a lot of tech debt. You know, like once you start generating enough, you see that it's just gone a bit more bloated or a lot more bloated. What is your observations here on how did you see this used to be for earlier models? Have newer models changed it? And is it still a problem? What are you doing about it? Yeah, I mean, once you unleash the agents on your codebase, there is some chaos. I mean, there's a saying like, let the chaos reign and then you reign in the chaos. That's a philosophy. So rather than you prevent the entropy, the entropy is going to increase of the codebase. Maybe your security posture also weakens because you're not reusing the same auth layers or something, like reusing stuff. Then you need to have agents which are self-maintaining that code, the janitorial, like the Roombas, right? So that's the approach we are taking. Like, let people ship quickly. Let some of the chaos reign. Maybe things are not scalable right now, but then you have the agents also cleaning up the mess, like all the way from flaky tests to bad abstractions to refactoring the code. For example, recently we ran agents to introduce Effect and XState into our codebase, so refactors constantly keep happening, right? So you only make sense of that mess once you actually have the mess there. Like then you figure out what's the best way. So I think both have to be like, it's like reigning in the chaos, as they say, and then reigning the chaos, right? Yeah, and I know you previously talked about things like de-slopping tests. Yeah. Can you tell me what is that? Is that a skill? Is that a— Oh, it's massive. Like, I'll tell you, our CI has gone through the roof. Like, I think we're probably spending as much on CI as tokens. And you're spending a lot on tokens. Yes, and also on CI. Like CI, I heard last time, like vendor quoted us like 13 million or something a year. We built our own by the way now. Yeah, but that would have been your run rate if you would have stayed on the vendor. That's right. So the thing is, like, these agents are now, because they're, like, writing code and CLI commands, they're driving a lot of compute. And all these code changes, you're still going to make them go through CI gates, your merge queues and everything, right? And you're, in fact, writing more linters, more simulations because you want to be sure. There are two problems here. One is, like, a lot of these tests are actually useless. So you already have agents which are de-slopping the tests, removing unit tests, doing more integration, end-to-end tests, redundant tests. Like, we have those running all the time. Now, at the same time, it's a constant battle. Like, your CI time goes up, but you want to keep them within three minutes, ideally. Like, that's how much—I think we are, like, roughly four minutes right now. And at the same time, then your sharding and all, like, the costs go up. And this is the reason, like, one of the things is, like, at our scale, like, it makes sense to just build our own. So we built our own CI, right, running. Right now what we are hitting is, like, compute capacity. Are we talking about CPU capacity? CPU, yes. Even GCP, unfortunately, has very little compute available, and Tropics took it all. So we were in the central region, and we couldn't even upgrade our database instance. We ran out of compute. And we had to, like— And you're, like, just to be clear, you're not the smallest of companies. I mean, you will not be a Google, but you have plenty of budget, you have good funding. Like, you're ready to give the money. Well, yeah, we have signed good contracts with them, but unfortunately they're all locked in by the large— because they're all running sandboxes, RL environments. I don't know what they're scaling. But they're also not just scaling GPUs anymore, but they're even the RAM CPUs. So, because I wrote about the CPU crunch in the primatic engineer a few months ago, but you're a good example where it's happening, right? It is happening in a big way. Right now we have a big CPU crunch as well, and it's just like a few companies shipping at that scale. I mean, for the rest of the world to get there, I think the biggest bottleneck is going to be the data centers, the energy, and those kind of things. The models are commoditized. Like, the models are there, but they drive a lot of that compute as they work. But now, a little bit talking about this CPU crunch. So, because I think this is— I'm super curious to learn more on both what you can do to get more CPUs and what you can do to maybe use your CPUs more efficiently, right? You mentioned Rust, and I wonder if this is about it, because Rust is as efficient as you can get your code down to, right? That's what we did. Like, I mean, we were on TypeScript earlier. Like, we just moved to Bun. I mean, that was just another— I was in Tahoe on a beach. I ran a goal, and over the weekend all the services were migrated to Bun. Dude, your idea of holiday. Yeah, I mean, I have the pictures. The team was so surprised. I just ran a goal command, and all the services moved to Bun over the weekend. So, just to understand, so you have this thing where you set up a goal, and then you do a long-running session in the cloud or on your local computer? Unfortunately, it's local. Like, I mean, we do have a cloud agent, but I'm the— I should be the one using it, but I have, like, a powerful machine, so I always go and— And then it goes off and it runs on your machine, and it runs a long-running task, and it usually completes it. It does. Like, I mean, it will sometimes, like, stall to ask a few questions. But some of these, like, interestingly, the distributed systems are now far easier to build. They don't need much hand-holding because there are only a few ways to get something right when it comes to the architecture. The product, it's like open-ended. Like, the UI you can build in 10 different ways, and that's where the taste, the judgment comes in a lot more. But we used to pay a lot for people doing complex scale distributed systems, things like Kubernetes and advanced caching, observability at scale, right? To be honest, those things are super cheap now. So you're saying that that part, that kind of this backend, even though they're complex, has been commoditized thanks to the models? Yeah, because they're like, as the saying goes, like, there is only maybe one way of building Unix. If you were to build it today, you're gonna come up with the same abstractions. Yeah, and I guess going back to, for example, Kubernetes, I had Cat Cosgrove on from the Kubernetes team, but she was saying how Kubernetes is, on one end, it's complex to onboard to because it's complicated, but they also have really good documentation. And I guess when you have something like this, and there's only, I guess, I assume there's only one or maybe two ways to do things, but when you have these agents are pretty good, when there's good documentation, no matter how complex it is, they can navigate the complexity. Those kind of things they're, like, great at. Like things like, okay, optimize my Dockerfiles. They can just keep running it against OrbStack locally and optimize it. Optimize my CI cache or do this branch migration. I mean, those kind of things are super easy right now. I mean, if you have good unit tests and testing, like, you can validate that, hey, things are coming up correctly and the smoke tests are passing, you can even try it on production. And usually things are working. It's not like we are shipping a lot more defects. Our defect rate, I would say, is going down. I mean, this might be a hot take, but, like, humans were writing slop, not AI. I mean, that's what I'm realizing. Like, AI is actually writing less slop at this point, if it can validate against something. So you're saying that in the past we shipped, like most teams would ship kind of slop or low-quality stuff when we didn't validate it. And to be honest, if we're just being honest, we didn't always validate it, right? We had more outages in the past. I mean, that's my observation. I still have to measure it, but I would say when we were shipping with hand. And you're talking about just cold rabbit, right? Now I'm seeing actually fewer outages. The more our codebase becomes agent-ready, the better our validation layers become, we are seeing less defects, not more actually, as the velocity is going up. So one more thing about quality is evals. Because of your business of offering a tool that gives you AI code review, model quality will be really important, and costs, and all of those things. How do you evaluate new model releases? And I'm obviously talking about the big ones, the OpenAI, the Anthropic, but also open models and so on. Because I assume it's kind of a big deal for you to know how good of this model is in terms of the quality, the cost, latency, and so on. That's right. I mean, right now we are also bottlenecked on our evals capacity. Like, oh, how so? I mean, the models, the cadence that they are coming out every few weeks now, because the RL thing has taken off, right? So people are coming up with back-to-back releases. The training runs are shorter and shorter, right? Which means that, I mean, every day we are, like, evaluating something. And they're like, of course, three or four classes of OpenAI models and Anthropic models, and we have, like, an ensemble of models in our product. So it's like, unlike other products, we don't let the user drop down and choose the model. Like, we do it for them. We know which sub-agents, which workloads, the model is going to be a good fit. So a lot of the time that our applied AI team spends is figuring out the price-to-performance of each model and the workload, to the point that we have zero bandwidth to even onboard another vendor. Like, we are a big token spender. We have been working with, like, even, you can see, the Gemini and—but we don't have bandwidth. That's been a big constraint, even with the open models. I know they're getting good enough. Each time we think that, okay, we could go use these open models, OpenAI brings down their models in price. Like Llama is a great model. Like, given the price point that that's being offered. Sounds like you're using it a lot. Yeah, we use a lot of Llama, like, in our product. We use other models. We use expensive ones for deep inference and reasoning, but a lot of code exploration, a lot of the context preparation is done by cheaper models these days, right? But the trick is, like, the open source is lucrative. The only thing it's helping is, like, putting downward pressure overall on the industry. We ourselves didn't get, like, much time to go and evaluate those. Do you build your own models? Do you either fine-tune or do some, like, post-tuning, or have you experimented with this? Early days we did, and that was an interesting realization that the frontier was moving faster than we could maintain those fine-tuned versions. Like, I mean, the shelf life of those investments were very short. So it was a little bit better for a short time, and then a new model came and boom, your advantage was gone. Then you're in a different zone altogether, right? So, and the other thing is, like, we are in the coding space where I think other people have plenty of data. That's what they are relying on. Yeah. I mean, if we, let's say, had really very different data stream, then maybe it makes sense, because the models are not good at trained on that. But if you look at coding code review, they're trained to some extent where the real trick comes in is in fine-tuning the harness. There are two parts of fine-tuning. One is the model. I think the bigger problem is, like, what's the right context? Because these reasoning models can go off track if they overthink. I mean, you might have noticed, extra, he's actually sometimes worse. It overthinks, right? Same thing. Like, I mean, depending on the context, we use a lot of, like, semantic analysis of the code, code graphs to make sure the token costs are, like, lower, while we also get very high-quality leads to go and track. So that's where, like, the effort goes in, is in fine-tuning the harness. And we have a very large open source user base, so we are, like, very popular in public repositories. Yeah, yeah. You have this thing where you make it really easy for open source projects to use CodeRabbit, right? Yeah. And a big reason for that was the feedback loop, because with open source, each review that happens every few seconds goes back into our pipeline to fine-tune the harness. So basically, just from a business perspective, I mean, you're losing money on every one of these because you're paying for it, but you're getting it because now you're able to make your product better, which hopefully it will be more valuable in the long term, right? That's right. And then this could be a win-win. Yeah. So the reason we have the price... This is, like, one of the very few use cases by the fintech industry where a specialized harness is still applicable. Like, everything else is getting eaten by general-purpose Claude Code. I mean, legal, finance, like, it's doing general-purpose work. But code review is one area where it's a breadth and a depth problem. You need a good recall, precision, and over there, like, a custom harness is very effective. And this is where, like, all this open source data is helping us tune that. I mean, I have a couple of agents just running, looking at open source all the time. One more thing about quality, as you mentioned briefly, formal methods. I'm interested, when did you start to experiment with it? You mentioned TLA+ and Lean. How is it going? Where are you using it? Because when I talked with Hillel Wayne, who was on the podcast, TLA expert, you all know him, he said that he thinks formal methods, especially TLA+, is probably an overkill for almost all software, even with AI, although he did say this about six months ago, so the progress was not there. I'm just really curious on where you're seeing success. So we started using it back in February or March when we built the Slack bot product. So that product had, like, some distributed states. Like, it had a queue, like, queue of the messages. You could steer the messages. It had... The complex states around, like shutting down the sandboxes. Like there was a lot of state machine there on the warm idle state and so on. So that's where that TLA+ models, they're not great for, like, things which are like algorithms. They're great when it comes to, like, messages and actors and interactions. And that's what, like, we used, because at 5.3 Codex wasn't as great as now Astra. Now I think it might make pure mistakes, but the way we made 5.3 Codex successful was first model it, make sure the state machines were correctly done before writing the code. And we did find issues in the initial implementation after we deployed TLA+, it could find all these, like, race conditions which we had overlooked. Yeah, simulations also help, by the way. So there were two techniques we invested back in February, March. We built, like, an entire, like, simulation harness where we could, like, fuzz and try, like, different interleavings. At the same time, we used TLA+ to model different states and explore those states and then find bugs. Nice. And I want to go back to the—not your commits, but the CodeRabbit commits. And this is the weekly commits, which are kind of similar to yours, except it's more granular. You can see. But there's a huge inflection point in August, very similar to how you went. It went from about less than a thousand commits per maybe, like, 500 or 300 in July to more than 5,000. We're talking weekly commits at the last data point. I wanted to ask, what about the cost of AI tools? You obviously have AI code review tools, you have AI generation tools. Where is it at right now, roughly? So our AI cost—I mean, mine alone would be 250K a month, 275K, something like that. Wow. You're really spending it. Yes. I mean, that's just me alone. And some of the people on the team are shipping close to me right now, right? And a lot of the also, like, the number two code committer is actually the CodeRabbit's background agent. So we also have, like, the agents looking at health and just creating PRs, and we just take them to completion. So we are spending, like, several million dollars a month. On tokens. Then we are also spending a lot on the CI/CD, right, and the sandbox environments. And it sounds like that was the CI/CD was on track to be a million a month, but you now brought it in-house, so you probably decreased it a little bit. Fifty percent roughly is what I'm hearing. I have to see, like, but we just shifted, like, over the weekend, this weekend, all to the local CI now. Okay, so for almost everyone else, this would absolutely freak them out, like having this thing, and they would be thinking, how can I rein it in? I also heard a lot of—I did research on startups and how their costs went, and for most of them, the costs spiked in August with Opus 5.5, and then it went down. Most companies are kind of just using it for coding. They're not running as many loops as you're doing. How are you thinking about the cost of AI, and why are you not nervous as a CEO and a founder? So I think you're going to use both kind of models. Like, so if you have a good model routing layer, then of course for some of the—most of the tasks, you can even use cheaper models. But some of the things you still have to bring in the expensive model. The other aspect is the fast mode, like you pay double with ultra fast. Those are really great, right? I mean, as a startup, you want to get there faster than the next competitor, right? Then the thing is, like, even though AI is getting commoditized, but it's now driving the second-order effect of the CI cost and the CPU crunch we are seeing, right? I think that's real. That's probably not going to go away unless, like, you are getting rid of unit tests or finding some other way of validation. Like some people are doing local validation and using that attestation to, like, pass the central CI. There could be some tricks, but the fact is, like, these software factories are going to be expensive. Right earlier, when I wrote that tweet, I know that I got a pushback on that tweet. We can talk about—let's maybe talk about this tweet. So this tweet, I'll also put it on the screen, but I'll read it to you. You quote tweeted this founder or engineer who shared how they're spending 32 million per year, and Claude spent building this massive multiplayer game. And you said that what we're seeing at CodeRabbit and elsewhere is this parabolic trajectory that even large seed rounds, like 8 million plus, are no longer enough to build a software business. Software factories need quality tokens at large scale compute to run CI sandboxes and other infrastructure to ship software at scale. At this point, some smaller startups exist only to funnel ideas to larger startups with more tokens. There's a lot of spicy stuff in here. But let's talk about—I know X is not necessarily great for capturing the nuance, but you had an observation there. You said that you're seeing this trend. Is it startups spending a lot on software factories? I think that's going to be inevitable. Multiple software businesses you can run, but if you're looking at hyper-growth, VC-funded startup in these competitive spaces, you will probably need more capital right now because now your bottlenecks have shifted. Earlier, you needed to pay mainly for the talent, so you'll get X Netflix, X Google, X Facebook engineers. They themselves will be founders. They will hire from the same pool. Give them nice talk, some really good laptops, and that's it. It was all the brain game. The IQ was the people, right? But you need to pay for them, right? In Silicon Valley, the base salary would have been 300K plus and stock and all that we're talking a million plus. Salary was the bigger piece in terms of running a business, in early stages. I'm talking about seed, Series A, even before you hire sales. A lot of it is the team, the dev team, right? Now what we're seeing is the IQs on the tap, the tokens. In fact, even the smartest people need to be driving tokens. And a single person like me, I can drive two or three million dollars of tokens a year, like blended models, not just always using the best model. And perhaps it should even be more once we have the factory concept taking off. So my point here was, like, the bottlenecks are now shifting in how much you can, like, drive these tokens, number of agents, the CI and the other infrastructure. Humans can make a lot more decisions in a day to drive millions of dollars of spend. Now you need, like, it's almost like marketing team, if you remember, like, I mean, 10 people can drive 100 million dollars of spend. Same thing we're going to see with engineers. Now a small team can drive 150 million dollars of spend or 30 million in that company, right? And that's how you stay competitive, because now you're going to compete with everyone who's writing software in an inefficient way, not in an economically viable way. If you're going to mass produce, you're going to see, like, companies, composite startups now spin up, which will be a brand, but with, like, many products they're going to mass produce, like just like how you do in a factory. I feel you're still a huge exception in everyone I talk to, including inside of Nvidia, there's a monthly budget of $10,000, and you can go above that, but most devs I talk to don't go above that. So I'll just play a little bit of devil's advocate with you. So you're spending a lot on AI, like way more than most. You're also a founder, so you have a really good idea of where the business is and what value it's adding. What value do you think you and people spending this much is adding in terms of, like, does this mean that you're clearly moving faster? But what is this resulting in? Are you getting more customers? Are customers spending more? Are you just, I don't know, getting more free users? You see what I mean? Like, because you're in a great position. Usually for an engineer, it will be maybe harder to connect, but you can clearly connect, like, what either you're seeing right now or what you're expecting to see as a result of this investment and spend. Yeah, all of the above. Like, one of the things you're going to see is, like, these point SaaS solutions, the days are over. Like now you're going to see really, like, companies doing a lot more software, right? And this is adding value in all parts of our stack. First of all, like, I mean, as I mentioned, like earlier I would, like, rely on vendors for some of the things. Now you can just build in-house much cheaper. There are many tools, like CI we moved in-house, we moved vector databases in-house, a bunch of this stuff. Those are very clear areas you can, like, go and build those differentiators quickly, right? And the other thing is even for the market, like the people have infinite appetite for software and these solutions, right? So there are companies which are gonna, like, for example, we are in the SDLC space, so all of our customers want to go and ship faster. Code review is not the only bottleneck they are running. Like now they are on the market looking for the next thing, which is, how do I solve my CI bottleneck? How do I do this? Like, so we're in that business. So we have to first of all go and learn quickly where the bottlenecks are gonna be. I mean, we're getting arrows in the back, front, whatever, right? But what that also puts us uniquely in a position is that we can offer these products faster to the customers. But even everyone wants to ship faster. Like that's the thing with the VC-funded companies, right? So the kind of, like, the range of tech you want to go quickly build, that's usually your moat. Like, I mean, you have like a few months. Now the problem with the software right now is, like, even like a lot of these startups, they used to have a moat of a few months before they got copied, but now they're getting copied within days and weeks. So you can't just go to the market in this competitive market with an idea which is shallow. Like, you get, like even Anthropic is gonna copy you in like two weeks right now. Oh, so I see where you're coming from, but I still want to push back on the fact that my sense is, okay, you're spending this crazy amount of money. You personally are on track to spend like two million per year, or let's just talk about per month. Like you and the team maybe are spending 500K in this month, maybe next month it'll be 600, maybe then 700. But in return, what you're getting is, you know, you are in the DevTools space with AI code review. You're now realizing, oh dang, we also have this bottleneck with CI. Maybe we'll do something with CI or help our customers do something with CI, and we do this because we're now so fast, we're seeing where they're gonna be in like six months or 12 months. So now we're now gonna start this product, we'll spend more. Then we realize maybe it'll be a feature flag, and we'll be able to roll out, and so on. And by doing this, you're almost like doing R&D for your business, and you're also building your products. And the bet is that you'll be expanding your market, you'll sell more to customers. But what I'm wondering is, are you not a little bit unique that this might be true for anyone? Not a little bit unique. That this might be true for anyone who's in the dev tools market, but for some company who's not. In any other market as well. So basically now you're moving in a compressed time frame. Let's say it took you like a while to go to market and do a PMF, six, nine months if you're a YC company, right? Now you can like go there and try that idea quickly. And if it's a shallow idea, now the risk is you can get copied. So basically you'll probably have to like do more, but you are now moving at a compressed time frame. The earlier the bottlenecks were in scaling the team. Like, so I needed to have the talent. You will like get all this VC money, scale the team. I don't need to scale the team at the same level, right? I mean, maybe the sales still need to scale because they have to do the meetings, but product, marketing, engineering, you can now do few people burning a lot more tokens, right? And these tokens spend right now, if you—and it's quality tokens, like you can get cheap tokens. It will take you longer to get there if you have the quality tokens. Your price per task will be high, but you get there faster at a high quality, right? Okay, so what I'm hearing is what you're actually saying with this tweet, which wasn't as nuanced, is if you are in the VC game, meaning you're raising venture capital because you want to build a billion-dollar company, which means you actually want to win your market. Like inside of dev tools, you might have a niche, may that be like AI code review or AI SRE or CI/CD, something like that. If you're starting today, you will probably need to raise a larger amount because you need to win this quickly. The game will be over pretty quickly, and you want to get meaningful traction before the big guys, Anthropic, OpenAI, can maybe copy you. Oh, and as an added benefit, you can have a small team to start with, and you don't need to worry too much about hiring. But by the time you would hire anyway, you will want to build some solutions and get customers and so on. So like everything is just compressed and you can pay more. Right, and the attention is compressed. Like if you don't get the mindshare after building your MVP in like maybe one or two months, that's the window you have now before someone just gobbles up that idea and runs with it. Okay, so that's pretty brutal. Yeah, I mean that's the situation right now we're dealing with. Like the window of opportunity to get the mindshare has become shorter. It used to be that you had smart people, they had built something. Right, and maybe it was a lot of product iterations. It was a SaaS product with a lot of, like, graft and care, or it was a complex backend system which just had a moat, like people couldn't replicate easily. So they would have, like, a shelf life of, I don't know, six to nine months before something, and they had that breathing room to get the mindshare. But now it's like a few weeks, unfortunately. And the company that's, like, spending a lot on tokens and all can get there faster. So you're going to see more companies like Rippling, which was a composite startup. You're going to see a lot more composite companies with these very efficient factories and the processes gobbling up a lot of that software market. You will still have, like, people tailoring at the last mile, but if you're buying these SaaS solutions and all, that's where you're going to see a big disruption. Okay, now I think this is fascinating insight. I mean, just KV, like, you're seeing this in Silicon Valley where everything is accelerated, always has been, compared to rest of the world. Let's say, you know, Germany will be always a lot slower, right? We've seen that with the industrial revolution, same thing happened. Like, I mean, one example is, like, for example, toys. Like an example, like earlier people used to make metal toys and very good craft. Even I grew up getting those Matchbox cars, like all metal, pretty good. Oh yeah, the old Matchbox, they were a lot heavier, right? I mean, when I was a kid, like they were very heavy, really good quality, but now you're seeing a lot of plastic. But mass-produced, the market has appetite for that. So yes, the quality is going to go down, right? But the companies that are going to invest in these factories, which need now actually as much investment as running a factory, it's almost like building a steel plant now, 50 million out of the gate, are going to mass-produce software. But just to push you a little bit more on that, because I really want to push you on this here, these software factories probably cost so much because there's not a turnkey solution. And let's say, given everything is so fast, in a year there might already be software factories you can buy, but you're saying that by that time, I mean, first of all, everyone can have access to it and your window of opportunity might have closed for a lot of businesses. Software factories are expensive right now, tokens, but then CPU, all the, like we talked about the CI, like you can drive a lot of spend just with the velocity alone. A single person cannot drive it, right? So the economics are going to force you into that factory model. It's almost like how... Like we've seen in the industry, even like, let's take another example, like watches. Like when the quartz watches came out, you could mass produce, and then you pay for Rolex, which is crafted. Now the question is, like, would artisanal software, which is very well crafted, people took a lot of time, will command a premium, or is the market just going to be okay with generic-looking, maybe slightly lower quality software, which I think is what the case is, and that software is going to be designed in a factory? Yeah, no, I see where you're coming from. I really want to disagree with you in a lot of ways, especially, but I see your point, and I think I have a hard time disagreeing with the fact that, look, if you do want to win a market right now, you want to give yourself the best time, the best opportunity, and if the theory is that to win, you want to be— Either first or fast and iterate and get it right, you have less time because an interesting thing that I'm also seeing, which is one more thing we could reflect on, is so far we've only talked about, we assume that your competition is either the big labs who, you know, see what everyone else is doing and they launch something, or the other startups who some of them raised more money and might be faster than you in their software factory. But the reality is everyone else is a competitor. Like Ramp launched their coding agent. Spotify is launching a platform for feature flags. So you have the existing tech companies who are also realizing, hang on, we built this tool inside because we're now big and we see and we think there might be an appetite. And what you said is every small company, like 50-person company like yourself, will need the tools that Netflix has. Netflix at some point might say, like, hang on, let's open source. So now I feel there's just a lot more competition for the same thing, like just coding harness. Again, what would have you thought that a credit card company would have a coding agent and people take it seriously? Right, and the thing is that whoever is not on this parabolic trajectory you just saw, they're going to get disrupted. A lot of people are not even ready for what we are seeing already. Like, their culture is not going to let them. So the way they are doing software development, like they have to rethink everything. Like, first of all, how many people they need. Like, people who are just good at writing code, they should probably look for other jobs if they don't know what to build. Like, people who are not software engineers, like artists, people who have taste, maybe should start coding. And then how do you allocate budgets? Like, you have to now really think about if you're hiring someone at like, I don't know, half a million a year, expect they're going to spend more than a million on tokens now. So earlier, like when you used to do this plan to show the VCs, you will have the salary as the main component, but now we're starting to show tokens, right? Yeah. Well, again, I just want to be clear, like this is for like a new startup VC building from zero, because for existing companies that have a business, often like physical or in the world, okay, they could be disrupted, but, you know, they have something stable to be there, and they're the companies who are spending the $200 max plan for everyone, right? I mean, if you have a hardware business and all, you probably have a better moat at this time, but if you're like just software... This software with very high margins. Those margins are going to compress, and you probably also have to, like, rethink. I mean, if you're not on the parabola, then companies coming from behind. Now, this exponential is scary. I mean, these kind of things start world wars, right? It's that scary. So that kind of companies are going to start coming from behind and start decimating some of these lower dinosaurs unless they go and transform themselves on this parabolic. And those companies, they are actually investing in factories more than talent. They're investing in just, I mean, paying for the best model possible. They're not cutting corners. They're just going fast. Let's talk a bit about hiring, because this is also a very, very interesting topic. First, can you tell me how did CodeRabbit grow over the years, and how has your hiring changed? Because you did mention that you're hiring a lot slower, right? In some areas, yes. In many areas, we are hiring rapidly. Like sales is one area where we have really doubled down on the field sales, because one of the things we found with generative AI is that the last mile is still a problem. Like people still need a lot of handholding in how to set up the systems in the right way. Yes. So those kind of things you can't accelerate with AI. You have to have those meetings. You have to have conversations. So those areas we are hiring rapidly. But other areas now we are, like, rethinking what we need to hire for. Well, let's talk about software engineering specifically. Software engineering is the most trickiest area right now. Every month we are learning something. Like one of the things we have found is, like, the problem-solving skills, system skills. Like people who came from, like, even I wrote a tweet about, like, I'm looking for people who have played strategy games. Let me read this tweet, because that was a really interesting one. So you are hiring, and you said, I mean, there's a caveat: We're hiring. The person who applies must meet one of the following two criteria to join: either has tastefully shipped at least 1,000 PRs in the last 30 days. Or has played at least 500 hours of simulation games such as Civilization, Age of Empires, Transport Tycoon, Factorio, etc. Almost like now, if you ask me what the software building feels like, it's like playing one of those games, right? Now, I mean, I talk about Civilization, like you basically make the moves, and each turn is like a release for me, right? And you learn something in the next turn. And while you are steering the civilization, you still have, I don't know, workers improving your roads and building fields and so on. That's how the agents are. And each turn, you are like just making these agents work on something meaningful, right, which we think is a great ROI for your civilization, right? Almost like now running these agents feels, to me, feels like building. If people have grown playing strategy games, I think they will be very successful with these agents right now. They can multitask. They have a range of things they can—I mean, at least they have a good problem list these agents can then work on. And how is hiring going for you so far? How easy or more like hard are you finding to hire these folks who have this kind of experience? So the way we are looking at it is like either someone who's already a tenured senior engineer at a good place, they have to come and first unlearn, right? Or you hire someone even who's like fresh out of college, but they have high agency but can unlearn faster. What we found that the people who are like coming with, like, who are not set in the ways and have come, especially from the background of like playing these strategy games and all, have been very successful. That's just my take on it. And that's what we are looking for. I think the human agency now matters a lot more than it ever did. A person who was 10x knew what to build is now suddenly becoming 100x with these agents, right? Because now you at least have a problem list or the most impactful areas these agents can work on, and you could just unleash these agents as much as possible in parallel. So we talk about individuals, but I know you have a concept called pods that you also call as war rooms. That's right. What are war rooms? So what we found is like that's another area, like I think we have to—like the collaboration has become even more important. So like in previous startups, run remote teams, we are building distributed systems, so it doesn't matter, like teams like in India or Eastern Europe, I've done all. But now, like what we're seeing is that now we're moving so fast. Like one person is on the steering wheel, but other people are shaping the software as it's getting built. And everyone is able to touch the software, by the way. Even product managers get to touch it, like, and the kind of PRs they're making is relevant to, like, change a copy on something or some information architecture flows, right? But everyone is, like, touching the software at the same time, and it's great if they can be, like, in a war room setting. So we kind of create these war rooms where we share, like, plans on a Slack canvas. We all agree on it, and we all collaborate in a room in a more face-to-face setting in real time. So you get everyone in a room, it's a table, it's all real time, kind of close the door and, like, all right, let's solve the problem, right? And we're crafting the software in real time. It's like sometimes you will, like, lay the foundation by running a goal, but when you're, like, at the stage where you want to, like, really craft the UI and the experience, you want to get people in the room and really craft it live. Yeah, it's funny. At Uber, we used to have up to 2017 something called the war room, which is like this: there's an important project, get everyone in, sometimes fly people in because they're distributed, and then spend maybe a week there, build up—this was pre-AI, so with AI it's a lot faster—but just build a proof of concept, whiteboard, code, and then have a prototype, and then everyone goes back to their teams and just kind of builds it from there. But it's now working, the core thing. Yeah, now we're doing that more often. Like, that's exactly the idea. And you need, like, people with different tastes. Like, you have product managers in the same room, you have designers, you have people who understand backend, the front end, and they are, like, crafting the software together. It is hard work, and sometimes it is late work. Because one of our first interactions on X was, I think I was writing something about how, like, teams are working really late up to the evening or into the night, and you replied, It's almost 2:30 a.m. and we're still shipping, and you're like, yeah, the kind of, you know, like crying smiley face, I'm not sure if it's good or bad. But can you tell me a bit on the pace? You talked about you're not sleeping as much before the recording. We talked about how, like, okay, at some point you need to get that under control, because clearly there's a time for excitement, but also when you're thinking of building something longer term, like you'll want to find the balance, right? How are you thinking about it? It's crazy right now. Like, I mean, it does, like, impact your health to some extent, because ever since these agents got better, as I said, like, there's, like, a huge opportunity cost if they are not— Not working all the time. And they do stall. They have questions. You have to, like, steer them. I read the reasoning traces. I guide them whenever I have to. And for that, like, I mean, I've almost, like, changed my sleep cycle. I mean, I take a nap. I wake up at, like, 3 a.m. to send, like, a few instructions because every hour, like, one or two agents are going to get stuck, and you have to go and, like, it's almost like Roombas getting stuck. You have to, like, unstuck them. That's what it feels like. But then that's kind of become a habit, but it's not sustainable. And we haven't, like, figured out the right interface to work with these agent fleet of agents at scale. Maybe it's really like a factory. Like, companies have to have a morning shift, night shift, because instead of owning the code, maybe you're now owning agents. During the day, I make the decisions, but I can be rest assured we are seeing the same session, the same agent, that other people can make decisions while I'm sleeping, right? So we are right now, because we're developing on our own machines, we are, like, bottlenecked on that collaboration right now. Yeah, the one thing I'll add is I wrote this book called Why We Sleep. I read it a while back, and it actually forces me to prioritize sleep because there's this thing with sleep, which there's some scientific evidence. I'm not an expert in this, but there's a suggestion that not getting enough sleep now, it's fine for tomorrow, and it's fine for next week, and it's fine for next month, but decades later, it can come back and bite you. And ever since, I don't know how scientifically sound it is, but just as when I was like, oh, it could happen, I'm like, okay, so, like, me getting a good night's sleep is not only important for me for tomorrow because I'll be, you know, like, you'll know, like, you won't do as well, but for my long-term health, like, I'm trying to hopefully try to do stuff that will make me more healthy when I'm 60 or 70, et cetera. So that's my only kind of, like, addition to, like, not sleeping. I would push a little bit, like, is it as important to wake up and steer the agent? Can it not wait until the morning? So I think it's more like a polyphasic sleep. Maybe I'm, like, exaggerating the health, but I think even if you look at how— But I think even if you look at how people used to work pre-industrial era, this 9 to 5 thing is a recent thing. And I mean, I come from, like, agriculture community, so people wake up 10 to livestock because you are, like, working based on other aspects, and you have to wake up and do that thing. Now it's the agents. You have to wake up to tend to that livestock, which is now agents. It's almost like farmville. I don't know whether that's the right analogy. So people used to have polyphasic sleep, and I'm doing that right now. Oh, you're right. Like, this was a thing in the Middle Ages. People would wake up in the middle of the night, and they would often, like, dine and stuff like that, and then go back to sleep. That's what I'm doing right now. But I will say the only reason I'm doing that right now is because I have to tend to the agents, which are on my local machine. Now, once you've really figured out the right abstraction of the factory, once you move to the cloud, then maybe it's going to be like a factory, like 9 to 5, morning shift, evening shift, or, like, round the clock, because these are not physical factories. You can have all of the Sun model. Like, we're still figuring out, but right now it looks more like polyphasic with this generation of agents, at least for me. Yeah. Let's talk a little bit about burnout. You wrote about burnout, and I'll quote you here. CodeRabbit is my third startup now, and I have worked just as hard for my previous two, NetSilo and FluxNinja. Burnout never happens due to hard work. It kicks in when your work does not lead to the desired outcome for a long time. On the family front, my two daughters were born during startups. They actually helped prevent burnouts. In just your personal experience, like, how have you thought about burnout and also just working so damn hard? Because it sounds like, you know, you're on third startup, like, you don't seem to be the—you're the intense type. Yeah. I mean, I really, when the second startup wasn't doing well, I mean, that's where I did feel dragged down, burnt out, because there was no outcome. That's the first time I felt the signs of a burnout. Because, yes, you were, like, working very hard, but without a very clear result or an outcome, just waiting. The anxiety kills you because nothing is happening. You're doing all this work, but nothing's happening, nothing's moving. So that's a bad scenario to be in. I think that's a status quo which I would never want to go back to. But for quota, like I'm working hard, but things are on the upside. Like we are, like, seeing the revenue grow like crazy. Our team is growing, the funding is coming in. It's always something happening. We are actually moving the needle as well. Customers are loving the product, the new capabilities. So even though I'm working hard, but I don't feel that I'm burning out or I have to stop at this point. And I think I can keep going on. And I wonder, like, I think in your case it makes a lot of sense because you were a founder in all the companies, so you immediately, you were connected to the outcome. Or, you know, for example, if the revenue goes up, you know this is just good because it will help your company grow or raise the next round or get a sale or whatever. You know, like, you know what can come. I'm wondering, just being empathetic for employees who are not founders themselves, they might have equity, but it will not be as big of a stake. Could it be that having autonomy, having visibility for the business, just knowing how their work is connected, that it might actually help? Because inside of people typically work hard and long, especially early stage. But I wonder if the autonomy here and seeing how your work is connected to the things that actually just helps, you know, get this kind of kick that you're getting, basically. Yeah, it motivates people. Like, first of all, we don't force people to work like I do. I mean, I lead by example. That's what I believe. And even if people can follow to some extent, it's great. But I do what I do. Like, I mean, rather than managing the people, I'm on the ground in the trenches with the developers. If there's incidents, I'm always, like, first there to help them out, no matter what the time. Yeah. And I think that does motivate people. They're all, like, happy that, okay, the leader is right. They feel, like, much more empathy to their life, right? Secondly, like, I found, like, the people who join startups are looking for this adventure. Like, it's a non-linear career trajectory. So they are leaving, like, very predictable salary, RSUs in some big company, right, to get, like, cheaper stock at startups, but then they are looking at the upside. And when that happens, it's, like, even more inspiring, right, when the companies are on the upward trajectory. So I would say, like, even the people who go and join the startups are looking for breaking that status quo, which... Things can be very mundane and boring in a larger setting. Yeah, and you're right. Like on a fast-growth startup, there are a lot of advantages. Like either, you know, if the fast-growth startup goes great and you have equity and there's an exit, you could have life-changing amount of money, meaning that you can buy a house, you could have education set up for your kids, and you now have the freedom to either take a risky startup or do a boring job, whatever you want. And if it goes like this and it drops, for example, Fast was a typical example in the, I think in 2022, which went bankrupt. You often, if the startup attracted really good people, those people all go to different companies because now their company shut down. But there's also often this network, this camaraderie that was there, and they often call each other years later. Some of them will end up at the next, you know, like really successful company. So I've actually changed my mind in, okay, the only downside, I guess, joining a startup is often you will need to work hard. It will often be in person, but typically when you're young and when, you know, this is not, you don't have other huge priorities in your life, it can be a huge career boost. It is. Like even financially, I wouldn't say it's crazy. Like that's how the Silicon Valley is. Like I think it does reward people who took that kind of a risk. I mean, there's no fatal failure somehow. Yeah. Like, and I've seen people who have just been through that experience have gained a lot in, like, even if it's like a bad lesson, at least they know what is not going to work. Yeah. And that's valued in whatever the next job they're going to go look out for. So you're kind of saying it might be more valuable sitting in one place, not doing much, not learning much, right? Yeah, that's where you don't grow. But if you go out and do this thing risky, I mean, there's no fatal failure there. Like, people love, in the Valley at least, people who have taken risk, whether you're founders or early employees, foolish enough to join. But at the same time, people reward that experience, right? And if it's successful, then it's like you're on an accelerated trajectory. And I think that's what I would encourage. Like even though, yes, I mean, it's becoming a little bit cost-prohibitive on the factory front, but still it's still the best time for people to go out and try this out. Like before all these, like factories become more established and start blocking some of these individuals. But right now, it's a very magical time right now. Like even a single person just driving these agents in the right way can create immense value. Now, I wanted to ask, looking back pre-AI to like your 25-ish years as a software engineer, as a founder, what are the things that you learned there or skills that you built that you think are really useful and helpful right now working with AI and in this, like, I guess, AI era? I mean, in terms of, like, entrepreneurship, it's always been like—I mean, again, it's almost like a video game. Like, I mean, you're playing all these, like, games where, yes, you have to think about, like, the product-market fit. I mean, all that is exciting. I think what I love is, like, creating value. Like, that's end of the day. Like, software just means to an end. Yes, you have to write code, and I enjoyed it, but it was, like, just means to an end for me. Like, the real thing was, like, our people. Like, it's really, like, really, like, making someone happy with the value you provide is always, like, the biggest reward, right? Everything is, like, then follows. Like, you will get the VC money, you will get the team, and people will join the mission. So it's almost like being a missionary in this kind of a world, like you're spreading a religion. And that's the fun part of doing a startup. Like, I mean, I've always been on the frontier. Like, I've seen, like, early telecom, I've seen SaaS, cloud takeoff, and each time it's always—what excites me is, like, okay, there's a lot of non-linear value going to be created at the frontier than at other places, and I always try to be at the frontier. What would you advise a software engineer who one day would like to work at CodeRabbit or at a place like CodeRabbit, just as a fast-moving place? What kind of skills should they invest in? What kind of practices, what kind of activities should they do outside of playing 500 hours of Factorio? Yeah, I think the agency is the number one thing I'm looking for right now when we interview someone, because right now, like, even if they're in a bigger company, because now a lot is happening on the side. If they're not doing side projects today, it's a red flag, because right now it's so easy to build that people who have agency are actually building something right now. They can hold themselves back, and that's always, like, a great sign. If you have side projects, I love it, right? And the second thing we are looking for is, like, people who have a range of experience. Now, for the depth, you have AI. It can go deep, drill infinite to performance, optimize something. But if you have, like, a range of experience... Now you can, like, build, let's say distributed systems. Let's say you came in from one space doing, let's say, observability databases. Now if you were to give agents, now you can do 10x of that. You're not just building that; you're building 10 other types of distributed systems. So the range also matters now. So those kind of things I'm looking for, right? I mean, it's like system thinking, whether they have range in skills, and high agency to keep these agents busy. As closing, what's a book or a paper you would recommend that you read and you found it interesting or fascinating? I think Living in a Black Swan Moment. I read this book. I mean, I'm not, like, spend a lot of time reading books these days. I've just been, like, building. But I think we are in that, like, the black swan zone, like what we are seeing right now. And what's the book? I mean, this was, like, The Black Swan book. I think I read—I forgot who the author was, but it was really, really... We'll put it in the show notes below. Yeah, yeah, yeah. But that was, like, pretty insightful, what we are right now going through is kind of a black swan moment. Like, if you look at five years back, I would not have predicted that AI would get so good that it's, first of all, like, driving my cars, it's driving the software, and it's, like, even some of the capabilities we are still, like, discovering with what these models are capable of doing, and we have not even hit the peak yet. So, yeah, so everything is going to change, and some of the things we can't even predict today, the way we're going to see the aftereffects. Well, Harjit, thank you. Yeah, thanks. Well, I cannot recall talking with anyone nearly as AI-pilled as Harjit is. To be clear, I was skeptical to start with when hearing about his $200,000 AI bill for a single month for a single person, and I remain pretty skeptical about several things. But Harjit explaining his thinking put a few things into perspective. Let's start with the part that I cannot agree with or recommend. Harchit's approach to waking up in the middle of the night to prompt his agents. He calls this polymorphic sleep. But to me, it sounds like deprioritizing sleep all to unblock agents. This feels very inefficient to me, and I would rather look at longer-running tasks in the cloud if you really want to be there to have your agents running at all the time, even when you sleep. But really, I don't see the obsession with agents always running. This is not a competition about utilizing your local laptop to always run at 100%. The part on how Harchit doesn't think a startup with $8 million of funding has enough money to build a software startup, I'm also skeptical. I get it that when Harchit assumes that all startups will send millions of AI like his startup does with 40 people, you burn through funding fast. I also get it that in dev tools, it can make business sense to build new products like a custom CI really fast, like CodeRabbit did. But outside of startups that sell AI dev tools, I just don't see it making sense to spend massively on AI just to be a little bit faster to market than your competition. The part I do get, though, is how much competition across AI dev tool startups is there today. It's not just CodeRabbit launching new products with their 40-person engineering team. It's also larger companies like Ramp or Shopify launching dev tools products, AI labs, and other large companies all trying to get the piece of the same market. So with this context, I do get why speed matters, especially for small startups, and why spending on AI to build stuff faster can make some sense. Finally, one practice I really liked is the concept of desk losing their driving license if they cause too many regressions. AI-pilled or not, I like this idea to force engineers to care more about reliability and shipping quality code. For more deep dives on software factories and engineering practices with AI, check out the link to Pragmatic Engineer Deep Dives below. If you've enjoyed this podcast, please do subscribe on your favorite podcast platform and on YouTube. A special thank you if you also leave a rating on the show. Thanks, and see you on the next one.