The failure mode of an unobserved agent is not a crash. It is a slow drift into being subtly, expensively wrong, which nobody notices until a customer escalates. By then you cannot reconstruct what happened, because you did not keep the trace.
The minimum viable instrumentation
- Full step traces: every tool call, retrieval, prompt version and model, with inputs and outputs retained.
- An offline eval suite that runs on every prompt change, including an adversarial set you keep adding to.
- Online evals on a sampled slice of live traffic, scored daily by a human for the first month.
- Per-request cost and latency caps, enforced at the runtime with a defined behaviour when they trip.
- A replay button. If you cannot re-run last Tuesday's decision exactly, you cannot debug it.
- Issues caught pre-customer
- 83%
- Mean time to diagnose
- < 10 min
- Rollback time
- seconds
- Cost variance
- ±6%