Back to catalog / Agent observability

Trace model, tool, and handoff spans in an agent run

Correlate model calls, tool execution, guardrails, and handoffs in one workflow timeline for diagnosis.

Goal

Identify where an agent task slowed, failed, or diverged from expectations.

Set tracing scope, retention, and sensitive fields that must never be recorded. [openai-tracing]
Wrap one or more Runner.run(...) calls in with trace("agent-workflow"): to group them in one trace. [openai-tracing]
Use SDK tracing or a custom processor to capture agent, model, tool, and handoff spans. [openai-tracing]
Call flush_traces() after the trace closes when immediate export is needed; configure a processor or disable tracing per policy. [openai-tracing]
Restrict export under organizational data policy; disable external tracing when prohibited. [openai-tracing]

A single workflow trace should locate tool and model stages; verify no keys or unnecessary personal data are recorded.

If spans are missing, check tracing configuration and processor flushing; do not log credentials to diagnose export.

Python 3.10+, openai-agents compatible with the pinned revision, and authorized model-provider credentials (except local mock tests)

Use a local or isolated test environment; use least privilege and approval for external data or side effects

Tracing is enabled by default and unavailable under ZDR. Keep tokens, personal data, and sensitive tool payloads out of trace metadata.

These values describe the upstream repository, not popularity or likes for an individual skill or playbook.

No metrics snapshot is available yet.

License: MIT