Agent evaluation needs operational traces

InfoQ's agent evaluation guide frames agent quality around task completion, tool failure recovery, memory drift, latency, cost per task, PII handling, and policy boundaries. Why it matters: Useful agent evals need to measure the whole system, not just final-answer quality. Trace-level checks expose whether tools failed silently, context drifted, or a task became too expensive to run repeatedly.

Try this: Add a small eval checklist to one agent task: success criteria, tool errors, retries, latency, estimated cost, memory changes, and any sensitive data in traces.