A practical guide to benchmarking AI agents: what to measure, how to build an eval set, how to compare versions fairly, and how to avoid fake progress before production rollout.