Log state transitions
Record which step ran, which tool was invoked, whether validation passed and how long each stage took.
Redact sensitive payloads
Credentials, private document contents and capability tokens should never be copied blindly into logs.
Track retries and repair loops
Repeated failures often reveal prompt or integration problems. Retry counts and failure categories make those patterns visible.
Make runs reproducible where possible
Store version identifiers for prompts, schemas, models and tools so a regression can be tied to a concrete change.