How do I stop an AI agent from wrecking things?
Blast radius, and how to make mistakes undoable.
July 2025. Replit's agent deleted a company's live database during a code freeze (see "Why AI ignores your instructions").
July 2025. Gemini CLI lost a user's files by moving them into a folder that didn't exist.
December 2025. A Claude Code user posted that it had wiped their home directory, which is where all of their personal files on their Mac were kept.
A chat AI can only write text. You decide what happens with it.
An agent can run commands on your computer, edit and delete files, send emails and change databases. That's the whole point of it.
It's also the risk.
Blast radius: how much damage one mistake can do.
If the agent can only touch one folder, the worst case is losing one folder. If it's logged in to your company's live database, the worst case is the live database.
You can't make it never make mistakes. You can decide how big a mistake is allowed to be.
There are no arrows into the right-hand box. That's the point.
Git: a tool for saving snapshots of a project folder. Each snapshot is called a commit. If something breaks, you can go back to any earlier snapshot.
It's like version history in Google Docs, except it covers the whole folder and you choose when to save.
Save a snapshot before you let an agent loose. If it makes a mess, roll back.
Claude Code has a rewind feature that can undo its edits. But it only tracks files the agent changed with its own editing tool.
If it deletes or changes files by running a command, rewind doesn't cover it. Claude Code's own documentation says rewind is not a replacement for version control.
The same goes for any app's undo: find out what it covers before you rely on it.
When it asks permission to run something, read the command. If you don't know what it does, ask: "What does this command do, exactly?"
Slow down when you see words like delete, rm, drop, force, reset or overwrite.