AI Operations and Incident Recovery

Trace the operation, not just the HTTP call

An agent request may include retrieval, multiple model calls, tools, and a later resume. Keep those steps linked to the user operation and release versions. Distinguish a provider error from an application decision to stop.

Monitor useful completion, quality checks, tool failures, repeated steps, latency, and cost. Redact personal data and secrets before exporting traces. More captured text is not automatically better observability.