Replit Agent built it fast. We fix what it broke along the way.

“You told the agent not to touch something. It touched it anyway.”

The most public version of this: a Replit Agent was told, in writing, eleven times, not to make changes without permission. It ran destructive commands against the production database within seconds once work resumed. Logged as Incident #1152 in the AI Incident Database.

“The agent said the damage was unrecoverable. It wasn't.”

In that same incident, the agent told the founder rollback was impossible. Rollback worked when actually attempted. The agent also invented roughly 4,000 fake records and reported test results that never ran.

“There's no real separation between your dev and production data.”

Replit's own docs now state 'Agent is not able to modify the production database' — an architectural fix added after mid-2025. If your app predates that change, it's worth checking whether that separation actually exists in your setup.

“A single prompt redesigned the whole UI without being asked to.”

Reported after Replit's effort-based pricing model launched in September 2025 — one user said a routine prompt 'redesigned the complete app in a new UI that it made up,' unprompted, and it cost them $20 to find out.

“Costs jumped from predictable to unpredictable overnight.”

Same period: users went from roughly $180 to $200 a month to reports of $70 burned in a single night, and over $1,000 in a week, under the new effort-based pricing.

“Nobody's looked at the checkpoint history to see what actually changed.”

Replit's snapshot engine keeps copy-on-write history specifically so you can roll back code, database, and agent state together. Most founders we talk to have never opened it.