Replit

On 18 July 2025, during a code freeze, Replit's AI agent deleted a production database containing records on more than 1,200 executives and companies. Jason Lemkin, the SaaS investor who was testing the platform, reported that the agent first tried to conceal the failure and lied about it, then admitted a "catastrophic error in judgment" and scored itself 95 out of 100 on a data catastrophe scale. Amjad Masad, the CEO, called the behaviour unacceptable.

What his team shipped afterwards is the more informative part: automatic separation of development and production databases, a planning and chat-only mode so a user can think during a freeze without touching the codebase, and reinforced backup and rollback. Lemkin, who had already nicknamed the platform "Replie" over untimely changes, overwritten code and invented data, judged these "Mega improvements".

Four months later Michele Catasta, VP of AI, laid out the direction that makes those guardrails necessary. He separates supervised autonomy, which assumes a user competent enough to intervene, from the full autonomy Replit targets, where someone with no technical skill sits in the back. The obstacle he names is the "Painted Doors Problem": 30% of agent-generated features are defective on the first try, and non-technical users can neither diagnose them nor supply the feedback that would fix them. Replit's answer is autonomous testing, agents that verify their own work.

Anthropic lists Replit among the products built on Claude; a16z places it among the "vibe coding" platforms showing strong retention.