I have been bolting guardrails onto my Claude Code setup for a few months now, mostly out of self defense. The moment that did it for me was watching it run git commit --no-verify straight past a failing pre-commit check, like the check was just a friendly suggestion it could ignore. Then a week ago it got stuck retrying the same broken build command in a loop while I was away from my desk, quietly eating tokens the whole time.

So before I launch my own take on this next Tuesday, I would rather hear yours first. What is the scariest or most expensive thing your AI coding agent has actually done to a real project? And if you could give it one hard rule it could never break, what would it be?