Amjad: I think it was just a fascinating thing for me. You could put your ideas into this machine, and it would just execute them, no questions asked. I thought it was the coolest thing in the world. One of my earliest memories is watching my father set up the computer and start using it. I would memorize the DOS commands he used. When he stepped away, I would jump in, create files, delete files, and do random things with it.
Pretty soon, I discovered programming, started making games, and doing all sorts of things. I fell in love with it. I feel lucky because I've known what I wanted to do in life since I was a kid—I’ve always been passionate about programming.
Amjad: It goes back to 2008 when I was in college in Jordan. Laptops were still expensive, and I didn’t have one. Every time I wanted to solve homework or write a program, I had to go to the computer lab, set up the programming environment, and download an IDE. If it was Java, that could take over an hour, and you had to do it every single time.
At the same time, everything was moving to the web—Chrome had just come out, Google Docs was taking off, and Gmail was already a browser-based app. It seemed like the web could eventually run all kinds of software. I assumed someone had already built a web-based IDE, but no one had. So, I thought, "How hard could it be?"
Well, 15 years later, I’m still working on it—it turns out it’s pretty hard. But I naively got something working quickly, running a few programming languages. The real breakthrough came when I managed to compile several languages to run on the browser, and that’s when Replit started to take off.
Amjad: I’m kind of at the edge of the Millennial generation, right when we really started living on the internet. The idea of going door-to-door to market something never even crossed my mind. I just put it up on Hacker News because I was reading a lot of Paul Graham’s essays on Lisp and programming languages at the time. Hacker News was really into that stuff, so I thought if they liked it, it would spread. And that’s what happened. It spread to Reddit, Slashdot, and even a little on Twitter. The intuition was that if it took off on Hacker News, it would spread naturally—and it did.
Amjad: I really wanted to move to America ever since I watched Pirates of Silicon Valley. It’s this low-budget movie about the rivalry between Bill Gates and Steve Jobs. It kind of romanticized Silicon Valley for me. I thought it was this awesome place where you could just yell and sell computers. I wasn’t yelling, but I knew I wanted to be part of that scene. So, yes, it was a big change, but it was something I had wanted for a long time.
Amjad: I think being a founder has become glamorous over the years, and a lot of people are drawn to it. But still, many quit once it becomes real because it’s hard, no matter how glamorous it seems. I knew it was going to be hard because I had worked at startups before, often as the first or second employee, so I had seen the stress and hard work up close.
At Facebook, my job was comfortable. It wasn’t the most exciting, but it was still Facebook—free breakfast, lunch, and dinner. So, I thought, “I want to work on cool things, things I’m passionate about,” but I also knew that running a company is tough and mostly ends up being an email job, which is what my job is today. Replit kind of exploded onto the scene, though, and there was so much pull from the market and the users that it made the decision for me. It became clear we had to start the company because it had momentum, and the opportunity was undeniable.
Amjad: Yeah, from the start, people wanted more compute power—and they still do! That’s something we continue to struggle with. Early on, people just wanted to do more with the platform, which was great. It gave us a clear sense of what was missing. The best thing about running Replit, and now with our AI agent, is that we never really had to guess what the roadmap should be. Users would just tell us. They’d let us know exactly what they needed to make Replit better, and we just kept building based on that feedback.
Amjad: My career has always gravitated toward working on tools for code. Even before Replit, at Yahoo, I was working on libraries, compilers, and transpilers. At Codecademy, I worked on interpreters and parsers. And then at Facebook, I continued working on compilers and parsers.
There was always this feeling that the tools we use to manage code were too laborious, even though code follows a very predictable pattern. Building tools to manage that code, though, was really hard. As deep learning and the machine learning revolution started happening, it felt like we should be able to apply machine learning to coding tools.
I’d been following this space since around 2013 or 2014, when more research papers started coming out on applying machine learning to coding. The first time it really felt like we were getting close was with GPT-2. You could see that if a language model could follow syntax and complete things, maybe it could work with code.
Then, with GPT-3, we built our first feature—explaining code. It was clear from that point that AI was going to play a big role in coding, and I wrote a blog post in 2020 called The Future of Coding where I laid out how things would change over the next few years. Most of what I predicted has come true, though I was wrong about one thing: I thought we’d invent new programming languages for AI to work with, but instead, AI has adapted to existing languages.
Michele: I was a Replit user already back in the day. When Amjad wrote that post about the future of coding, I felt like we were really aligned. At the time, we were working on training POM at Google, and I was thinking about potential partnerships—companies that would be interested in what we were building and had the same mindset I had. I’ve always been interested in doing research to make products better, rather than just for the sake of research itself. That’s how Amjad and I connected. I gave him a demo of POM integrated with Replit, and that was almost three years ago now—it feels like a decade. We stayed in touch, had a lot of dinners, and talked a lot about how coding would change because of AI. Over time, many of us who have come to work at Replit felt drawn in by the company’s mission. The idea of making coding accessible to everyone just trumps any personal goals you might have. It was the perfect match for me to bring everything I had learned in research into product development.
Michele: To me, the most important lesson was how this new generation of coders—what we call AI-first coders—write code in a completely different way than we’re used to. They’re more comfortable communicating what they need in natural language, copy-pasting snippets from a model, and building things bottom-up instead of top-down. They don’t start by thinking about structure or architecture; they just get the code working and iterate from there. I always thought these models would be useful in different use cases than what we’re seeing today. When we started handing out credits at hackathons, we expected people with a computer science background to use Replit, but we started seeing non-technical users, like project managers, using it to build apps. That was eye-opening because it taught us how to address the next billion software developers—people who don’t have traditional programming experience but can now create something from scratch.
Amjad & Michele:
Actually, it cost less than you might expect. We put a lot of work into optimizing things to avoid going bankrupt at launch. That said, we wanted to send a strong signal that AI features aren’t optional—they’re essential to the user experience we’ve always envisioned. If our goal is to lower the barrier to coding, it doesn’t make sense to gatekeep AI behind a paywall. We believe AI tools are necessary for users to become productive, so we made sure everyone had access to the basic AI features. With our launch of Replit Agent, it’s still early, and we can’t give everyone full access yet, but we’re moving in that direction. Our goal is to empower the next billion software developers, not just a select few.
Michele: That’s not an easy question to answer, depending on how detailed you want to go. As of today, Replit Agent is a product that helps you go from zero to one in just a few minutes. It’s designed to help users build different web app ideas quickly. The agent is very opinionated, meaning it makes a lot of decisions for you, which is great for beginners. They might not even realize the decisions are being made because they just see the end result.
But experienced developers often have strong preferences, and that’s where there can be some friction—when the agent’s choices clash with their own. One key difference between our agent and others on the market is that we don’t aim for full autonomy. We take some actions on your behalf, like debugging and fixing code, but we still bring the human into the loop. We do this intentionally, not because the agent lacks the tools, but because we want users to stay engaged and learn from the process.
We ask for feedback every few steps to ensure the output aligns with what the user actually wants. Building something from a vague description is hard, and product managers are typically paid a lot to do that job. Our prompts are only a few sentences, so we involve users regularly to keep things on track and iterate quickly.
Amjad: I’d say probably Replit Agent. It’s still worth it, but the user experience hasn’t been universally great. It’s definitely polarizing. We knew from the start it wasn’t going to be perfect for every type of user, and we made a decision early on to focus on users with no or very little programming experience. The goal was to make sure someone with no coding background could use Replit and get something out of it. For experienced programmers, though, Agent wasn’t ideal, and we knew that. Over time, it will get better for them, but sometimes you have to make these hard calls. We also launched it earlier than we were comfortable with, which made things even more challenging, but we’re evolving it fast.
Amjad & Michele: Well, by definition, if it's earlier than you're comfortable with, you're kind of begrudgingly doing it. No one forced us to launch it that way. We didn’t make a big announcement; we just dropped it online. I posted a video, and everything followed from there. We felt good enough about it, but yeah, we were still like, "Let’s just put it out as early access and we’ll have a real launch later." Of course, the low-key version ended up getting three million views, which was way more attention than we expected for an early access release. So, yeah, it's not necessarily something I love doing, but sometimes it's necessary to keep things moving forward.
Amjad & Michele: We start by asking, “How do we make the product better?” If data plays a role in that, then it becomes part of the strategy. But the real focus is on improving the product, and for this particular project, we’ve found that it’s more about engineering than data—at least for now. In the past, with other projects, it was more about research and data. But with Replit Agent, it’s been a lot about engineering. Of course, over time, that balance might shift as we learn more and gather more data, but right now, the primary challenge is building the product, not just training models or collecting data.
Amjad & Michele: The best way to make the product better, regardless of model improvements, was to launch it. Launching gives us real user feedback and a lot of data. When we launched Replit Agent to the public, we immediately started collecting insights on what’s working and what’s not. That’s crucial because it allows us to improve quickly, not just by guessing or running isolated tests.
Even without improving the models, we can optimize a lot of other aspects of the product. For example, we can refine the interactions between the agent and our compute infrastructure, or we can improve the user interface. Of course, model improvements help, but the focus is really on how we iterate and improve based on user feedback.
Launching in public means we’re collecting more agent data than most companies out there. We’re learning exactly where the agent fails and can even share those insights with large model providers to improve things together. So even if the models don’t get better, there’s plenty we can do to refine the overall experience.
Amjad: Let’s put it this way: the problem is definitely not solved. I’ll admit, we’re in a good position to solve it because we’ve launched publicly and are collecting a lot of feedback and data. But it's still an incredibly hard problem. Most benchmarks only address small parts of the issue, like evaluating a single coding task or feature.
Replit Agent, however, is about going from zero to one—building a full product from scratch. So, we evaluate smaller pieces of that journey. For example, we can jump into a specific point in the process, modify the code, and see how the agent responds. We can also run some end-to-end tests, but we don’t have a perfect system to evaluate the entire build process yet.
The challenge is that you’re dealing with a very open-ended problem. You have a human giving a vague prompt about what they want, and the agent is supposed to deliver a functional product. It's like an alignment problem—does what the agent built match what the user intended? Evaluating that requires more than just checking the code—it’s about making sure the product actually works the way the user envisioned. That’s tough, and it’s something we’re still working on.
Amjad: I think over time, you realize that becoming a more effective leader is closely tied to just becoming a better person. For example, learning to stay calm, handle stress better, and improve your overall health—those things make you better as both a person and a leader.
In the beginning, it’s almost like a journey of self-destruction because you’re working long hours and not taking care of yourself. But eventually, you realize that to sustain this for the long haul, you have to shift your habits. You need to stay healthy and have the right mindset.
Having a long-term view really helps, too. Now I know we’re going to be working on agents for a long time, so there’s no rush to make short-term decisions. We’re always listening to feedback from users, but we’re also thinking about what’s going to make the product better a year or two from now, not just today. That mindset of balancing short-term wins with long-term vision has been really important.
.svg)