Comment by ElFitz

These kinds of situations are why I gave my AI agents stray thoughts (automated insights / suggestions from a separate llm call with some curated context) that trigger on loop / rabbit hole detection.

Quite a bit of false positives, but it hasn’t had any ill-effect so far. Aside from increased quota usage.