STANDARD · Profile revision 1

OpenTelemetry GenAI Semantic Conventions

Development-status OpenTelemetry semantics for generative AI agents, model calls, retrieval, tools, plans, evaluations, and MCP activity.

standardopentelemetrygenaiagentsmcp

Why it is in the map

It is the broadest current vendor-neutral GenAI span vocabulary in the reviewed set and materially defines model, agent, tool, plan, evaluation, and MCP evidence.

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

The convention documents logical agent identity, version, conversation correlation, and parent-child span structure.

Tracevity interpretation

These fields can support agent-definition and execution correlation, but they do not by themselves identify an authenticated initiating principal.

Limit

A stable ephemeral runtime instance and a uniform parent/subagent relationship beyond span ancestry are not fully specified.

Primary evidence (1)
02

Principal and delegation

Not documented

The current accepted convention does not define principal and delegation evidence.

Tracevity interpretation

Agent identity and span ancestry must not be represented as human or service identity, delegated authority, or credential provenance.

Limit

A deployment may add custom attributes or correlate identity-system records, but that behavior is outside the reviewed convention.

Primary evidence (1)
03

Instruction and context

Partially documented

Instruction and selected-context artifacts are representable when capture is enabled.

Tracevity interpretation

A trace may preserve useful instructions and retrieved inputs, but only the emitted subset is reconstructable.

Limit

The convention does not establish a uniform context fingerprint, source-version guarantee, or explicit missing-context representation.

Primary evidence (1)
04

Decision artifacts

Partially documented

The convention defines recorded plan and evaluator artifacts in bounded circumstances.

Tracevity interpretation

An emitted plan or evaluation is observable evidence, not hidden chain-of-thought, a true motive, or a complete policy-decision record.

Limit

Selected alternatives, general decision summaries, policy outcomes, confidence, and refusal reasons are not comprehensively standardized.

Primary evidence (1)
05

Model activity

Documented

The convention provides broad semantic coverage for model request and response activity.

Tracevity interpretation

The vocabulary supports detailed model reconstruction when an instrumentor emits the relevant optional and required attributes.

Limit

Actual field presence depends on instrumentor and capture configuration; proxies can make provider attribution differ from the upstream provider.

Primary evidence (1)
06

Tool and MCP activity

Partially documented

The convention documents detailed tool-call and emerging MCP span semantics.

Tracevity interpretation

Emitted spans can reconstruct calls and results, while downstream correlation and retry representation still depend on instrumentation boundaries.

Limit

Tool arguments/results are often opt-in and sensitive; the MCP portion remains Development status.

Primary evidence (1)
07

Effects

Not documented

External effect evidence is not standardized separately from tool execution and result data.

Tracevity interpretation

A tool request or successful result cannot be promoted to proof that an external state changed.

Limit

Destination logs or independently observed state would need separate correlation outside this convention.

Primary evidence (1)
08

Human control

Not documented

The current accepted convention does not define complete human-control evidence.

Tracevity interpretation

Approval, rejection, edit, interruption, takeover, and rollback request must remain absent unless another correlated source records them.

Limit

The open proposal may change; this observation is bound to the observed date and current accepted specification.

Primary evidence (1)
09

Outcome

Partially documented

Execution-reported success and failure artifacts are representable.

Tracevity interpretation

These fields establish reported execution outcomes, not independently verified external success, rollback, compensation, or resulting state.

Limit

Partial and ambiguous downstream outcomes require application-specific evidence beyond these span fields.

Primary evidence (1)
10

Trace integrity

Not documented

End-to-end trace integrity properties are not part of the current GenAI semantic convention.

Tracevity interpretation

Trace IDs, span parentage, and timestamps are correlation artifacts, not integrity guarantees.

Limit

Collectors or backends may add controls, but those controls are not established by this standard profile.

Primary evidence (1)
11

Portability

Partially documented

OpenTelemetry-native transport and correlation portability are documented.

Tracevity interpretation

Wire portability does not ensure every destination preserves Development-status GenAI meaning without loss.

Limit

Import destinations, conversions, and field-level semantic-loss mappings are not comprehensively documented here.

Primary evidence (1)
12

Privacy

Partially documented

Capture-time privacy cautions and selective content recording are documented.

Tracevity interpretation

Opt-in content capture can reduce exposure but does not establish how a backend retains, deletes, hashes, or controls the data.

Limit

Retention duration, deletion, screenshot/code handling, and customer-controlled archival remain implementation-specific.

Primary evidence (1)

Related compatibility evidence

Directional mappings involving this system

These mappings preserve their own format/version, evidence, and fixture scope. They do not change the documentation posture in the twelve findings above.

Exercised August 29, 2026

OpenTelemetry GenAI development snapshot → OTLP 1.11

Representative GenAI agent, model, and tool attributes survive local HTTP/protobuf serialization, while principal, approval, destination state, and complete-capture claims remain outside the representation.

OpenTelemetry GenAI development snapshotcommit 67dff024110be5bd9f318006e733f4078e0f4c97
OTLP 1.11.0 HTTP/protobuf1.11.0
Read the mapping evidence

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: The convention documents logical agent identity, version, conversation correlation, and parent-child span structure.
  • Context: Instruction and selected-context artifacts are representable when capture is enabled.
  • Decisions: The convention defines recorded plan and evaluator artifacts in bounded circumstances.
  • Models: The convention provides broad semantic coverage for model request and response activity.
  • Tools: The convention documents detailed tool-call and emerging MCP span semantics.

Explicit gaps or uncertainty

  • Principal: A deployment may add custom attributes or correlate identity-system records, but that behavior is outside the reviewed convention.
  • Effects: Destination logs or independently observed state would need separate correlation outside this convention.
  • Human control: The open proposal may change; this observation is bound to the observed date and current accepted specification.
  • Integrity: Collectors or backends may add controls, but those controls are not established by this standard profile.

Source register

7 reviewed primary sources

  1. Semantic conventions for generative AI systemsOpenTelemetry project / CNCF · STANDARD SPECIFICATION · observed August 28, 2026 · current
    Open source ↗
  2. Semantic Conventions for GenAI agent and framework spansOpenTelemetry project / CNCF · STANDARD SPECIFICATION · observed August 28, 2026 · current
    Open source ↗
  3. Semantic conventions for generative client AI spansOpenTelemetry project / CNCF · STANDARD SPECIFICATION · observed August 28, 2026 · current
    Open source ↗
  4. Semantic conventions for GenAI eventsOpenTelemetry project / CNCF · STANDARD SPECIFICATION · observed August 28, 2026 · current
    Open source ↗
  5. MCP span modelOpenTelemetry project / CNCF · STANDARD SPECIFICATION · observed August 28, 2026 · current
    Open source ↗
  6. OpenTelemetry Semantic Conventions 1.44.0OpenTelemetry project / CNCF · STANDARD SPECIFICATION · observed August 28, 2026 · current
    Open source ↗
  7. Issue 95: MCP tool-call approval semantic conventionsOpenTelemetry project / CNCF · OFFICIAL REPOSITORY · observed August 28, 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