What every Medium/High-risk agent must track in production. Standard health-check signals (uptime / errors / latency) are necessary but not sufficient — they tell you whether the agent is running, not whether it is doing what it should.
- Output consistency / distribution shift. Compare current output patterns to baseline (Population Stability Index, KL-divergence, or simpler equivalents). Establish the baseline at go-live, not later.
- Escalation rate. How often the agent hands off to a human. Falling toward zero is a red flag, not a success — it can mean the agent is overreaching, not that it has become reliable.
- Decision audit trails capturing business logic. Not just tokens and latency. For each decision: what data was considered, what policy governed it, was the policy enforced, who reviewed the outcome. This is what auditors actually ask for.
- Cost per execution at the step level. Token cost spikes at individual workflow steps reveal infinite reasoning loops, unexpected tool invocations, behavioral anomalies. Cost monitoring is a governance signal, not just a finance signal.
- Exception routing patterns. Pre-computed routing rules built into orchestration. Patterns of exceptions reveal where the workflow itself is wrong, where the data is dirty, or where the agent is being abused.
These run continuously, not on a quarterly audit cadence. A misconfigured agent can ship thousands of bad outputs between point-in-time audits.