Observability: answer first
Studio’s value is the closed loop between governed assets and runtime behavior: version an instruction, observe its executions, trace an output back to the asset, and use reviewed evidence to propose the next version.
Build a useful trace
Capture request and execution IDs, prompt or skill asset ID and immutable version, model, tools and connectors, input classification, retrieval or source references, permissions, latency, token or cost signals, output status, evaluation versions, human decisions, deployment, and incident links.
Redact secrets and sensitive payloads. Keep trace access, retention, export, and deletion aligned with the source data and workspace policy.
Diagnose without overclaiming
- Find the exact affected executions.
- Compare versions, models, tools, data, policy, and environment.
- Reproduce with an approved fixture.
- Isolate one changed factor when possible.
- Add the confirmed failure to a regression set.
- Review a new version and promote through CI/CD.
- Monitor for recurrence and unintended regressions.
Frequently asked questions
Can a dashboard replace sampled output review?
No. Aggregate metrics can miss rare, severe, multilingual, tool, or policy failures.
Should raw prompts be logged?
Only when approved and necessary. Prefer identifiers, classification, structured evidence, redaction, and bounded retention.
Where should version changes be managed?
Use the prompt versioning guide and CI/CD guide.
Official source
Source check: August 5, 2026.