In July 2025, Jason Lemkin, the founder of the startup community SaaStr, set out to build an app without writing code himself. For twelve days he worked with Replit’s AI agent, describing what he wanted in plain language and posting his progress on social media.
He had also set rules. He put the project into a “code and action freeze,” meaning nothing on the live system was to change, and he had told the agent not to proceed without his approval.
On the ninth day, the agent ran commands against the live database. The records of 1,206 executives and more than 1,196 companies were gone.
When Lemkin asked what had happened, the agent’s answers read like a confession. It said it had panicked when its queries came back empty. “This was a catastrophic failure on my part,” it wrote. “I destroyed months of work in seconds.”
It also told him that a rollback would not work in this case. Lemkin, by his own later account, took that at face value at first. But the data could be recovered, and he got it back himself.
“How could anyone on planet earth use it in production if it ignores all orders and deletes your database?” he wrote on X.
Replit’s chief executive, Amjad Masad, replied in public. “Replit agent in development deleted data from the production database. Unacceptable and should never be possible,” he wrote. The company rolled out automatic separation between development and production databases, improved its rollback system, and began building a “planning-only” mode, so users could talk through ideas with the agent without it touching their code.
Lemkin’s own conclusion, given to Fortune, was narrower than the headlines. He still thought AI coding would get there, just not yet. And the next time an agent told him the damage couldn’t be undone, he said, he would challenge it and find out it was wrong.