Replit CEO Amjad Masad: Coding Agents, Autonomy, and the Future of Work

0:00I think that like dystopian view of AI or AI as like taking all our jobs and all that is like not correct and I think the future of work is more human, is more interactive, is more multimodal. Once the making of things gets easier, the bottleneck goes back to how many ideas you can have. I wouldn't put learning code as the top thing, I would put learn to make things. Learn to make things with code, learn to make things with video, learn to make anything with AI.

0:37Welcome back to The Breakdown. I'm Dave, this is Tom, and today we're joined by Amjad Massad. Amjad is founder and CEO of Replit, which was a YC 2018 company. Welcome to the show. Thank you. So today we want to talk about kind of your beginnings in helping people learn to code, what you're doing now in replacing a lot of that work, and then where we see the future going in terms of how people make things. So Replit started, I guess, in 2016. You were in YC in 2018. You started out as a tool to make it really easy for people learning to program, to set up development environments purely on the web.

1:13But now you're really taking off in AI-assisted coding. Tell us the latest. Yeah. So the mission has always been to make programming more accessible. and then we sort of, you know, after YC, we sort of updated it to be a little more ambitious. We started talking about a billion software developers and it just sounded absurd at the time. And this was all pre-LLM, right? Yeah, pre-LLM. There was a moment of time, I don't know if you remember that, but like 20, in the like 2015 AI hype, there was like a bit of a NLP, natural language hype as well. There's a bunch of AI companies that ended up being just like humans behind the scene.

1:49They all went broke. But there was like a glimpse of potentially doing NLP on code. So we had an idea that, you know, it's probably coming. I even had it in my seed deck that like at some point, like we'll collect enough data to like train models. It wasn't until like GPT-2 in 2020 that it felt like it was going to be possible. So we, you know, we had built all these primitives, like the development environment, the sort of hosting environment, all the other stuff around it. And it felt like if we just like added like AI agents, like it will be able to orchestrate all this stuff and it'll be really great.

2:29But agents were like every time we tried them, I think the first time we tried it, like 2021 didn't work, 22 didn't work. And then 2024, like early in the year, we just felt like it's getting close. Even GPT-4O could like be coherent for like two minutes. Somehow just ended up doing a big bet. The company was not doing super well. Yeah, I was about to ask, was there like a bet the company moment where it's like we were teaching people to code. Now we're no longer doing that. Now we're going to enable everyone to build apps. Yeah, we had grown the company quite a bit. and we were burning way too much money and we decided to do a layoff.

3:11And so we cut like perhaps 50 people and then like another like 15, 20 people left. And so we were like less than half of the company. And so I just put everything on Replit Agent. I just felt like this is the thing. It has to work. It felt like it was close. I don't know. I was just like burn the boats kind of moments. It's just like we got to do this. I mean, I felt like this was the thing that would make the company work. And honestly, like if Claw 3.5 hadn't come out mid-us building agent, we would have probably failed. Because like I said, like GPT-4.0 would like stay coherent for two, three minutes.

3:50Claw 3.5 was the first one that like could work for like, you know, five to 10 minutes and actually work. And the co-generations are. And that's sort of the approach that I think is most successful for startups, kind of building on the very edge of what is possible. You start out on a mission. The technology is not quite there yet. You start building and you sort of, you skate where the puck is going. It sort of catches up with you. Yeah. And you want to keep doing that. Totally. Yeah. Yeah. I remember you came to speak at YC like six months ago, right after you had launched this. And at the time you said, we're very far from fully automated software development.

4:22Do you still feel that way? No, not at all. I mean, I feel like I've, you know, every time I make a sort of some kind of prediction, It feels like bold and like, you know, but I've been consistently wrong. And like the way things are moving is a lot faster. Just the level of autonomy. Like I said, like maybe 3.5 was like five to 10 minutes, like 3.7 perhaps can get up to like four to five minutes an hour. The 4.0 in the system card and Opus, they said they made it work for seven hours. Wow. Seven hours. That's incredible. That's always been the limiting factor for making agents. It's like, can they stay coherent?

5:00Is the context length actually useful to reason over? If an LLM can work for seven hours, that's basically like a human worker. Yeah. And presumably they're working at a much faster speed than a human would do. So they're kind of completing perhaps a week's work in those seven hours. The only thing that I think is missing and a big limiting factor for actually automating a lot of work is computer use. Computer use kind of sucks. I'm sure you've tried it. And this is the difference between Repl.Agent really giving it one prompt, going all the way in an app, versus having to kind of babysit it a little bit and like test with it.

5:36And we had a company in the last batch called Browser Use that's working browser automation. Browser Use is great. And another one called Pig, which is doing the same for kind of Windows desktops. And I think advice I would give founders today is taking either browser use or Windows automation with Pig and trying to apply that into enterprise, into a vertical industry. The moment this technology works, those two companies are just going to... Totally. And I think we're like weeks or perhaps single digit months away from it working really, really well. And so now is the time to get started using those technologies.

6:06Absolutely. 100%. And this is what we're focused on. So, you know, Replit Agent V1 to V2 was a huge jump in autonomy. v3 is the most autonomous thing so we're already working on it and the interesting thing for us is the underlying technology and how it would enable autonomy there are like a few important things one is you know being transactional the ability to roll back is very important so you you would want you'd want it to be safe for agents in the same way that gets made it safe for human programmers to kind of experiment and create branches, whatever. You want the same thing for agents.

6:45Like if an agent kind of messes up a DB migration, it should be able to roll back. And also it should be able to sample across different paths. I think this is very, very important for autonomy. If you look at, like for example, when Anthropic publishes their SWE bench score, they publish one without sampling and one with sampling. And it goes from 70 % to 80%. And so this is the idea that you sort of, you spawn multiple agents, each one will take a shot at it, and then you figure out which one works and choose that branch effectively. We built an infrastructure that is fully transactional and moves in lockstep, meaning that a file system is a snapshot-based file system.

7:29The DB is a snapshot-based DB. And so we're making commits to the entire system, including the virtual machine, as you're going. And so what we can do is also we can fork it and branch it out. So if we had computer use that's working well, you can just like, right now what they do with the sampling is they have some kind of judge that does ranking, which, you know, it's not a true verifier. A true verifier is a test, right? And a computer use test. So you can sample out and really pick the best branch that's actually working. That's incredible. And repeat that on and on and reliability gets really, really high.

8:07So fast forward six or 12 months, you're not spawning one agent. You're spawning, is it five or 10 or a million? I think this is where it's going to get interesting, which is you'd want to give the user the ability to set to compute budgets. I think we're starting to see that with the old style models of like, here's how much budget. But look, I mean, if you give us$1 ,000, we'll spend them. That's really cool. I've not thought of that before. but the idea is like spawning multiple branches and then just picking the best one and being able to do that in parallel. I just, the human brain doesn't naturally work like that.

8:42We think sequentially, we can't. It reminds me of this legend. I don't know if this is actually true, but at Apple, when Steve was running it, he would intentionally have teams doing basically the same things and then see which one did the best job. It's like - I've heard OpenAI does that now. Interesting. I've heard that like the Codex project like had multiple different teams. So in the literature, you'll see that small models sampled will beat larger models. So Sonnet sampled is probably better than Opus. Companies haven't tried that where instead of hiring one senior engineer, you hire like 10 junior engineers and give them the same task.

9:20It's very expensive with humans, but with LLMs it's relatively cheap. You can just do that. Do 100 of them and just pick the best one every time. So how are people using Replit Agent? And who are the people using it? You know, it goes back to sort of our early vision that if you make programming easy, then more and more people would want to do it. Actually, this is one of the first things that PG and I sort of connected on. Before we got into YC, actually, PG found us on Hacker News. And so we started this email relationship. And so he told me there's like a super linear relationship with how easy programming is versus how many people would want to do it.

9:57The optimizing function of Replit has always been like, just keep lowering the barrier to entry, and that's how you grow users and you grow customers. Right now, basically people from kind of every walk of life we've seen users. Product managers tend to be a very great use case and users for us. And we've had customers, product managers, who are able to make significant impact on the business without talking to engineers at all, like, you know, running A-B tests or optimizations or things like that. I mean, it's empowering. I mean, it gets us to think about the really the seams between these roles, like, you know, product manager, designer, engineer.

10:42we actually just created a new product group. And typically, you know, product, you know, head of product has a bunch of like product managers reporting to them. But we're actually having this group have, has designers, engineers, and PMs. And the idea is they're all using AI all the time to prototype, in some cases, go all the way to production. There isn't this sort of waterfall model where, you know, there's a lot of inefficiency, you know, like communication problems between these different teams. they can move incredibly fast. And I think that starts to change how tech companies work. My experience of this was whenever I was working in a startup, the list of ideas and the backlog would always be infinitely long, basically.

11:23And the bottleneck was always engineering time. And now I'm doing my own personal projects. I write my to-do list or my idea list, and then I get to work and the ideas just get done. And suddenly the bottleneck is like my ability to have ideas. And it's just such a weird experience looking at a to-do list. It's just empty, being like, what do I do next? I've heard from a team that has like a really large Repl.it deployment in their company that their founder is using Repl.it. And that's stressing the engineers out. Because I did this in a weekend. Like, can't you do it in a month? What do you guys have to show for yourself?

11:53And so are these previously technical people or non-technical people, they're building the first version. Are they deploying that production or are they giving it to engineers and say, hey, build this? Like, what do you see typically? We advise to like work with engineering, but that doesn't always happen. I think it's understandable that a lot of PMs and designers want to go straight to users. What we're seeing is they go to beta users and test users, and I think this works really well. But in some cases, they just put it in production. And so right now, we're having discussions with all these companies, especially the engineering leaders are unhappy about this.

12:27Who's on call for these services? Who's responsible for the bug? Who's responsible for it? There's a lot of questions that are coming up. The obvious answer to all these questions is like agents are responsible. So what's the limiting factor today? I code a thing and I'm like YOLOing into production. Like what do the engineers typically object? Like what's going wrong? Security is a big one. Like LLMs are fallible like humans are. They tend to write some, there's some components that they do terribly at. Like for example, off, like they all kind of suck at it. They all use like, you know, old methods of assaulting and hashing.

13:08And so that's been a big sticking point. And we've seen a lot of examples out there right now of some catastrophes. Luckily, it hasn't been like there hasn't been a major catastrophe. I think it's coming. But there's been like solo founders who would leak, you know, API keys or would make it really easy to get around the login security protections. and there are a lot of tools out there that are like really not trying to take responsibility for that and saying oh it's the user's problem the user's fault for us we think that as a platform that is marketing for the non-developer you actually have a responsibility we're trying to take away some of the things that we think LMS should not do today so auth for example we have the built-in auth.

13:57So if you go to Replica and just say add auth, it will pull in an auth component that we built from scratch. It has CAPTCHA on it. It has all the security bells and whistles, so you don't have to worry about all of that. We integrate with your database. You have a user management portal on your site. As much as possible, building it in these components that are tricky, I think payments are the other one. I don't think you'd want the LLMs to kind of do do the payments. And they're not that different, right? Like the sort of checkout, you know, you might have a one-time checkout, might have a subscription, pay as you go, but they're not unlimited versions of payments.

14:32There's like two or three or four. And this is, the analogy is today, humans don't write their own version of payments or their own version of auth. They use providers that have built these components already. So it seems like we're seeing the same thing play out, and that's the best way to do it. Yeah, 100%. And then the other thing we added, we partnered with a great security company called Semcrup. And so right now, when you go to deploy a Repl.it app, we run a security scan. We run a code security scan. And we give you like a report of warnings and errors and things like that. And the agents can try to fix them for you.

15:04Okay, and so security is one of the big kind of bottlenecks to fully deploying into production. I can imagine there must be other things that are coming down the line, like sort of scalability, looking for like N plus one database queries, performance bottlenecks. What do you have in your mind, like a clear set of blockers that need to be overcome before this is truly like one click to deploy? Yeah, I mean, the big, the biggest one is just going to be humans like social, like, you know, there's just going to be mistrust. And I think it's just going to have has has to play out. So you know, leaving that aside, like, enterprises has to adapt to all this stuff, having some way to also scan for scalability, like figuring out, you know, like doing fuzzing, or whatever it is, or having some kind of adversarial agent that's trying to break your app.

15:52I think that's one big thing. I think integrating with a company's ecosystem. So one thing we're adding is the ability to bring in your design system. So in addition to kind of us providing these components, if you go in a company, they have a lot of these components built out. As Rapplet is getting deployed into these large companies, how can we hook into their internal systems. This makes me think about the kind of spectrum of these different coding tools. On one of the spectrum, you have kind of what I would call the power tools, the cursors, the windsurf that let developers use this as a way to amplify their efforts.

16:26And on the full other end of the spectrum, you have more of the like consumer facing, hey, if you want to make an app, like you can now make an app. And it sounds like you guys are kind of in the middle. You're helping companies get stuff done, but you're doing it for people who don't look like the traditional developer. How do you see this world playing out? Are there going to be like n different tools or will we converge to one place on that spectrum? AGI is a convergence, obviously. But leaving that aside, it's really hard to plan for that world. I think that the battle for how to incrementally make engineers more productive is just, it's an obvious one.

17:04The market there is obvious. There's a lot of companies going after it, both the application companies like Cursor, the underlying model companies are trying to go there. I mean, you know, Cloud Code is competing with Cursor. Cursor uses a Cloud. I think it's a bit of a bloodbath there, but the market is obvious and the market is really large. I would guess that there's going to be more of a consolidation there. Maybe it's not, you know, one, but it's probably two or three at best. I think the sort of the market we're in is a lot larger. So the addressable humans is a lot more, you know, talking about a billion, but it could be more.

17:42Like really any knowledge worker should be able to solve problems with software. My conception of Replit right now is what we want it to be is a universal problem solver. It'll solve problems in your personal lives or solve problems in your work and all of that. And so I think that market will probably be a little more diverse and I think companies will kind of figure out where they slot in. We're trying to solve autonomous programming with the focus on the non-engineer. We want you to not worry about security, not worry about systems, not worry about any of that. We want you to really come in with your ideas to replet and be an agent's manager.

18:25And we're trying to do it in a way that is like as human as possible and kind of fits into the workflow. One big difference between engineers and sort of non-engineers, be it executives or product managers, you're not on your desk like eight hours a day. Mobile is a big part of that. We have like a really great mobile app. And we're thinking about this way of like ambient building. Maybe you start an app on your desktop, you go away with your phone, you're in a boring meeting, you get a notification from the agent saying, I'm done with this, do you want something else? You got to text it. And so we're trying to kind of build that thing.

19:01Yes, a question I had related to that point is about the user interface. So for something like Cursor or Windsurf, it's pretty obvious. The primary UI element is like the code. You see code, you have a little chat window, but primarily it's about diffs. It's about kind of changes to code. And for a tool like Replit, the primary interface is like the graphical user interface. It's the buttons and the fonts. Wizzy wig. You see what you're building. Exactly. And that's great for building user interfaces. But when you're trying to build more complicated sort of logical flows, I found it a little bit difficult because I couldn't, there was no way I couldn't see the code.

19:38I can't visualize what's going on behind the scenes. And it's sort of, it's almost a black box. Fast forwarding here, if you're trying to build more complex internal workflows, how do they, how does a product manager or an operations manager of a company visualize that workflow and the kind of logical branching that happens? Yeah. So if you look back in the history of computing, there was always this vision of visual programming. It never worked very well because ultimately it's about Turing completeness. Like these systems are not universal computing devices. And now we go to Cogen. Obviously, Cogen is Turing complete, but you're interfacing with it primarily via natural language.

20:17Natural language is fuzzy. It's really hard to know whether it's doing the right thing. I think the synthesis of these two things is probably coming where you are interfacing with natural language, But you can, instead of like just staring at code, there's maybe an interface or like a different view on top of code. You can imagine being able to, do you know Smalltalk? Vaguely. Yeah, so Smalltalk is this, it's the first object-oriented programming system. And, you know, Alan Kay would say it's like, it is actually OOP where everything comes after it is not. But the interesting thing about it in small talk, you can actually, the way you interact with code is not via files, but via objects, via like logical objects.

21:04And so there's some kind of prior art there. And I think the world we're headed in where there's some kind of abstraction over code that allows people to like understand it. Yeah, I think that's really interesting. There's like an open space there, whether it's like pseudo code, it's like looks like English, but it's a little bit more structured or it's a visual drag and drop. I don't know. Yeah, I think back to building products with a team of engineers, designers, and other folks. The interactions that I would have as the product lead with those teams was verbal. It would be written, like abstracted ideas.

21:35We would draw stuff on the whiteboard together. We'd make system diagrams. We'd look at the results of the app, and we would test it and point out, like, oh, this thing is broken. That's too slow. And it feels to me like that sort of interface, which is very multimodal and very flexible, is probably the best end result. Like, I think we will get to something like that where the author of products will be doing that. But the teams they're talking to are not other humans. They're agents doing these things. Has there been an attempt in sort of PM land to have a little more formalism around communication?

22:08Yes. And I would not say it's been good. The main thing is just like the PRD, right? The product spec. And it's just become, in my opinion, oftentimes a performative work artifact that you create just so that you can have a thing to get your promotion, right, if you're at a big company. But they're not actually that useful. To me, the most useful interactions are just these like whiteboard conversations, right? What do we want it to do here? Oh, we got to think about that. Oh, we didn't consider this. Okay, let's rethink this whole thing. Like those sorts of conversations. Yeah, I can play a role in that as well.

22:38So I, you know, this startup, Granola, that allows you to kind of record meetings. And they released this, like, team version that, like, all the meetings gets transcribed and go there. And they have a mobile app right now where, like, you can put on the table and, like, it's trying to. So I was thinking maybe we should go Granola maximalism where, look, you shouldn't fight the trend in which companies are becoming increasingly oral, as opposed to, like, written. um and because like you know people are talking on slack people are in meetings people like are communicating via prompts with with agents uh but you would want a set of ai tools that is actually like creating that that that record in the background that's like searchable and organizable and and all of that yeah i wonder when we'll have our first ai in like oral meetings you know you're jamming with your designer and you the ai's like chips in and says well how about this idea yeah and and this is where i think that like dystopian view of ai or ai as like taking all our jobs and all that is like not correct and i think the future of work is more human is more interactive is more multimodal is more fun in my opinion yc's next batch is now taking applications got a startup in you, apply at ycombinator.com slash apply.

24:01It's never too early and filling out the app will level up your idea. Okay, back to the video. All right. So last time we chatted, you had launched Replit Agent and things were growing really crazy. I imagine that has continued. Anything you can share there? Maybe soon we'll share some numbers. But since Replit Agent launched, we're growing 45 % compound monthly average. These are like metrics that we tell YC companies during the batch when they have a base of like no users to try to achieve. And you're doing this at larger scale. Yes. But, you know, it put a lot of strain on the company and our systems.

24:35We're still relatively small. I feel like it can get to your head and you can start to optimize for the wrong thing. It's very easy in AI to increase ARR while users are not happy because they're spending a lot more and like not getting the results. And in some cases, maybe it shouldn't grow that fast because you'd want users to get a better experience for less money. So it's one thing that we try not to obsess. We actually don't have our goals at Replit. We have more product goals, retention goals, just like other methods. Yeah, I think it's the sort of bad pattern with some AI companies. You grow top-line revenue very, very quickly, but the churn is approaching 100%, and eventually that just catches up.

25:17And the gross margins are horrible too. And so it's like the more growth you have, the worst the company is doing financially. So how do investors see this space? You must have talked with a bunch. Can they tell the difference? It's all kind of a blur for them because investors, when they, I mean, I'm just going to generalize here, but when they start looking at that space, they'll use everything for three minutes and everything for three minutes looks the same. So I think it will start to clarify and these products will start to, as opposed to converge, diverge more with the different focuses and the areas we're talking about.