Most teams running agents have monitoring (latency, tokens, error rate). About half have evaluation (known questions, agreed answers, run before and after every change). Monitoring tells you the agent ran. Only evaluation tells you it reasoned correctly.