At what point do you clear context instead of letting an agent try another quick patch?

I spent an hour yesterday watching an agent make four micro-fixes in a row only to land on the exact error code we started with.

It gets caught in this endless trial-and-error loop where it just keeps tweaking lines of code without checking root cause.

I started forcing myself to kill the session as soon as I see a second repeated error message.

How do you break that loop once the thread gets too messy? Do you restart from scratch or change up how you feed it error details?