Agent Observability

If you can't replay a failed trace, you can't fix the agent.

Logging "model returned X" is not observability. An agent has structure — graphs, loops, tool calls, retries, sub-agents — and the trace has to reflect that structure or the failure is opaque.

When to use this skill

• The user is moving an agent from notebook to service.
• The user reports flaky or intermittent failures with no reproducer.
• The user is adopting OpenTelemetry, LangSmith, Langfuse, Arize, or similar.
• The user wants to run evals against real production traffic, not just a fixed set.

What a complete trace looks like

A single user-facing request produces one trace, with nested spans for every step. Each span has: