Start locally

RUNTIME FRAMEWORK · Profile revision 2

OpenAI Agents SDK tracing

Built-in tracing for OpenAI Agents SDK workflows, including model calls, tools, handoffs, guardrails, and custom spans.

vendorruntimeopenaiagents-sdktracing

Why it is in the map

It is a high-signal vendor runtime trace system with documented workflow-level visibility across model, tool, handoff, and guardrail activity.

Tracevity boundary

A documented field can support reconstruction, but its presence alone does not prove completeness, authenticity, authorization, or external settlement.

Profile findings

What the documentation can—and cannot—establish

Each row distinguishes source evidence from Tracevity's bounded interpretation. “Not documented” describes the reviewed sources; it does not prove an implementation cannot record the artifact.

01

System identity

Partially documented

Workflow/run grouping and agent transitions are documented trace features.

Tracevity interpretation

This supports execution grouping but does not establish an authenticated principal or a complete stable agent identity tuple.

Limit

Stable agent-instance ID and agent-definition version fields are not established by the reviewed page.

Primary evidence (1)
02

Principal and delegation

Not documented

Principal and delegation evidence is not documented in the reviewed trace overview.

Tracevity interpretation

An agent handoff is orchestration, not proof of human or service authority delegation.

Limit

Separate platform or application identity records may exist but are outside this profile's evidence.

Primary evidence (1)
03

Instruction and context

Partially documented

Prompt evidence is documented as part of trace-driven debugging.

Tracevity interpretation

Some instruction evidence may survive, but the reviewed page does not establish complete context capture.

Limit

User/system separation, source identifiers, retrieval context, fingerprints, and missing-context representation are not fully specified.

Primary evidence (1)
04

Decision artifacts

Partially documented

Recorded guardrail activity is documented as a control artifact.

Tracevity interpretation

A guardrail result does not reveal hidden chain-of-thought or true internal motive.

Limit

Policy schema, evaluator detail, refusal reasons, and other decision artifacts remain incomplete or unknown.

Primary evidence (1)
05

Model activity

Partially documented

Model-call presence in the built-in trace is documented.

Tracevity interpretation

This establishes observable model activity but not every desired field in Tracevity's model-activity family.

Limit

The reviewed overview does not establish exact provider/version, complete request/response, token, latency, error, or cache fields.

Primary evidence (1)
06

Tool and MCP activity

Partially documented

Core tool-call/output and agent handoff visibility is documented.

Tracevity interpretation

Tool activity is observable, but complete call IDs, retries, server identity, latency, and downstream correlation are not established here.

Limit

Field-level export semantics require more specific documentation or implementation testing.

Primary evidence (1)
07

Effects

Not documented

External effect evidence is not documented as distinct from tool output.

Tracevity interpretation

A tool output must not be presented as proof that the requested external state changed.

Limit

Destination-system evidence would need separate correlation.

Primary evidence (1)
08

Human control

Partially documented

Approval support is documented at the workflow/runtime level.

Tracevity interpretation

The exact recorded evidence for approval, rejection, edit, interruption, takeover, and rollback request remains unproven.

Limit

Trace field names, actor attribution, timestamps, and rejection/intervention semantics are not established here.

Primary evidence (1)
09

Outcome

Partially documented

Trace-reported workflow and tool outcomes are documented.

Tracevity interpretation

Recorded outputs do not independently verify external success or resulting state.

Limit

Partial, ambiguous, rollback, and compensation semantics are not established by the reviewed overview.

Primary evidence (1)
10

Trace integrity

Not documented

End-to-end trace integrity properties are not documented.

Tracevity interpretation

Server-side trace storage or dashboard visibility must not be described as an integrity guarantee.

Limit

Storage and custody controls require separate official evidence or testing.

Primary evidence (1)
11

Portability

Partially documented

Default OpenAI export plus additional or replacement trace processors provide documented export-control surfaces.

Tracevity interpretation

A custom processor can receive SDK trace/span objects and can replace the default exporter, but that is not a standardized interchange guarantee.

Limit

No OTLP/OpenInference schema, lossless conversion, destination acceptance, or complete-buffer delivery guarantee is established by the reviewed processor API.

Primary evidence (2)
12

Privacy

Partially documented

Sensitive model/tool content capture and controls to exclude it are documented for current Python and JavaScript SDK tracing.

Tracevity interpretation

Structural tracing can remain enabled while sensitive generation/function payloads are excluded; this is capture control, not a complete retention or deletion policy.

Limit

Sensitive capture defaults on. Retention duration, deletion, hashing, screenshots, and downstream processor behavior remain incomplete or outside the reviewed trace contract.

Primary evidence (2)

Reconstruction reading

Do not collapse these findings into one score.

This profile describes documented evidence surfaces. Whether they are sufficient depends on the reconstruction question and the other identity, authorization, tool, and destination records available.

Potentially useful evidence

  • Identity: Workflow/run grouping and agent transitions are documented trace features.
  • Context: Prompt evidence is documented as part of trace-driven debugging.
  • Decisions: Recorded guardrail activity is documented as a control artifact.
  • Models: Model-call presence in the built-in trace is documented.
  • Tools: Core tool-call/output and agent handoff visibility is documented.

Explicit gaps or uncertainty

  • Principal: Separate platform or application identity records may exist but are outside this profile's evidence.
  • Effects: Destination-system evidence would need separate correlation.
  • Integrity: Storage and custody controls require separate official evidence or testing.

Source register

5 reviewed primary sources

  1. Integrations and observabilityOpenAI · OFFICIAL DOCUMENTATION · observed August 28, 2026 · current
    Open source ↗
  2. Agents SDKOpenAI · OFFICIAL DOCUMENTATION · observed August 28, 2026 · current
    Open source ↗
  3. Data controls in the OpenAI platformOpenAI · OFFICIAL POLICY · observed August 28, 2026 · current
    Open source ↗
  4. OpenAI Agents SDK Python tracing v0.22.0OpenAI · OFFICIAL REPOSITORY · observed August 29, 2026 · current
    Open source ↗
  5. OpenAI Agents SDK JavaScript tracingOpenAI · OFFICIAL DOCUMENTATION · observed August 29, 2026 · current
    Open source ↗

Next question

What evidence does your use case require?

Move from documented field presence to an explicit reconstruction target, or examine selected directional format mappings without changing this system's documentation posture.

Trace Reconstruction Requirements Explore compatibility evidence