Summary
Y Combinator’s Paul Graham sits down with Replit CEO Amjad Masad to discuss how Replit helped pioneer vibe coding and where AI-driven software creation is heading. Masad recounts how watching users paste code between ChatGPT and Replit, copying errors back and forth, convinced the team to automate the whole loop.
The conversation covers early technical challenges: running models to completion until a program actually works, building a robust edit tool when models hallucinated line numbers, and handling models that got lazy mid-function. Masad says the agent harness in 2024 was about ten times more complex than today’s because models were less capable, which means products are constantly commoditized and teams must keep building the next thing. He also explains why Replit’s existing infrastructure for databases and deployment gave it an edge, and how the company reorganized around a new agent team.
Graham and Masad also discuss why early users tolerated tools that failed around 30% of the time, how it feels to live in the future, and Masad’s view that Replit is now a platform for self-driving companies rather than just vibe coding. Essential viewing for founders and developers following AI coding tools.
📺 Source: Y Combinator · Published November 07, 2026
🏷️ Format: Interview
