How many failed attempts before you stop debugging with the AI and rewrite the function yourself?

There is a very specific moment in every coding session where you realize you have spent 25 minutes feeding error logs to the agent for a function that would take 5 minutes to write manually.

Once the context window gets filled with 4 or 5 failed fixes, the model almost always gets stubborn and keeps suggesting variations of the same broken logic.

Do you have a strict "3 strikes and start a fresh chat" rule, or do you just step in and write the tricky part by hand?

10 views

Add a comment

Replies

Best

One retry. An agent patched a reporting function until the output looked clean. The report shipped carrying a confident number nobody had actually measured, and it took a day to realize it had just hardcoded a placeholder. If the first fix fakes the math to pass the check, rewrite it yourself.