Tell Replit what you want in English and it writes the code, provisions a database and ships the app. Replit chief executive Amjad Masad told Paul Graham on Y Combinator's How To Build The Future that the loop runs until the program works, not for one chat turn.
Graham hated it.
Masad warned him, "Don't look at the code," because the output was machine code for humans to ignore. Masad said Replit was already generating 50 to 60 percent of user code with autocomplete before full agents, so the team knew the rest was possible. He got the design from watching beginners copy code from ChatGPT into Replit, paste the error back, and repeat by hand until something ran.
Replit automated that repair loop and taught models to provision infrastructure with tool calls for databases, Redis and hosting. Masad said the model labs asked how Replit made it work, and the 2024 harness was 10 times more complex than today's because weaker models needed robust edit tools and retries around hallucinated line numbers and lazy functions.
That history is why Masad calls Replit an iceberg, with a simple prompt box above the water and sandboxing, deployment and a custom transactional file system below it. The collaborative editor helped, because an agent could join as another user and the IDE team could shift to building edit tools for agents. Enterprises now care about that depth for reliability and security, while Masad said some rivals ship insecure agent-written apps, and his guardrails can make Replit slower and more expensive.
The price gap is what forces a budget rethink, with early Replit users choosing it over $100,000 outsourced builds for about $1,000, per Masad. Coverage of his enterprise push reports banks and manufacturers replacing projects once scoped at $10 million with $250,000 of Replit, reported by Whatfinger and not detailed in the interview. It still failed about 30 percent of the time in beta, so teams with strict uptime needs should check deploy logs and permissions before they delete the outsourcing line item.
