Y Combinator · Nov 7, 2026 · 44m · 36m saved
With Amjad Masad, Co-founder and CEO of Replit
Photo via repl.it
Replit's autocomplete was already writing about half of its users' code before the company shipped an AI agent in 2024. Amjad Masad, co-founder and CEO of Replit, told Paul Graham, Y Combinator's co-founder, on its How to Build the Future series that the agent was the next step in a long climb, and that the "vibe coding" label no longer describes what Replit sells.
"We don't think we describe ourselves as vibe coding now. We think Replit is the platform for the transformation to self-driving companies." — Amjad Masad
Amjad has been building toward a browser-based coding tool since 2008, and he says years of database, hosting and file-system work are what let Replit's agent run real infrastructure while newer rivals ship insecure apps.
The full interview is covered here so you can skip it. 44 minutes of audio, 8 minutes of reading.
Here are the 7 takeaways that matter.
Key Takeaways
Replit's agent grew out of autocomplete that was already writing 50–60% of users' code
Non-coders were pasting code and errors between ChatGPT and Replit by hand
Replit's moat is infrastructure built years earlier: its own databases, Redis hosting, deployment and a reversible file system
Enterprise customers are replacing $10M projects with $250K of Replit
Adoption usually starts with internal tools and spreads
Amjad says companies should measure the return on AI tokens instead of rationing them, because one person can now drive 100x the next
Replit now pitches the "self-driving company," with itself as customer zero
Most banks are still using AI the way they did in 2023, Amjad says, about two years behind Silicon Valley
An hour on Replit flipped a telecom's staff from 80% anxious to 80% optimistic
1. The Agent Grew Gradually
Amjad rejected the idea that Replit's 2024 agent was a sudden leap. By then the company's autocomplete was already producing a large share of the code its users wrote.
Replit's autocomplete was writing half its users' code before the agent shipped
Amjad's estimate of that share: "You could probably say like 50, 60%." The trigger was watching non-coders in the office stitch two tools together by hand. They generated code in ChatGPT, ran it in Replit, and carried every error back for a fix.
"They would copy the code, put it into Replit. Replit would produce some error. They'd copy the error, put it into ChatGPT... and then copy the fix, put it in. It was such a manual process." — Amjad Masad
Amjad concluded the loop could be automated into a single agent that runs until the job is done, instead of stopping after one turn the way most tools did then. Replit demoed the agent in July 2024 and launched it that September. Because the models were far weaker, its harness was "10 times more complex" than today's, and the model companies themselves were caught off guard that it worked.
"They actually reached out to us and they were like, 'We're so surprised that Replit works. How did you make it work?'" — Amjad Masad
