A good fit for
- → Production LLM debugging
- → Prompt/version comparisons
- → Cost and latency analysis across agent steps
Know the limits
- → A trace is evidence of execution, not proof of output quality.
- → Unredacted traces can become a second sensitive database.
- → SDK and self-hosted server compatibility must be checked.
Your first useful workflow
Choose the trace boundary
Use one trace per user-visible job and child spans for retrieval and tools.
Add meaningful metadata
Record model ID, prompt version, outcome and billing units.
Create a regression set
Promote privacy-reviewed failures into reusable evaluation cases.
The practical take
Instrument the backend. Tracing secrets never belong in a browser bundle.
Sources & evidence
Capabilities above are summarized from these official documents. Workflow suggestions are editorial guidance, not a claim of measured product superiority.