OpenTelemetry GenAI Semantic Conventions
Development-status OpenTelemetry semantics for generative AI agents, model calls, retrieval, tools, plans, evaluations, and MCP activity.
1 documented families4 gaps or conflicts
Public beta · Trace Map + local Inspect + Gate
A trace tells you what was recorded. Tracevity tells you what you can reconstruct.
Independent, source-linked intelligence for understanding the identity, authority, context, actions, effects, integrity, portability, and privacy evidence in AI-agent traces.
Documentation observed through August 29, 2026 · 0 Tracevity-tested system profiles
TRACE / 00:19:42.031DOCUMENTATION VIEW
principal?initiator not documented
agentA-117instance correlated
tool callT-093arguments + result
external effectUNVERIFIEDdestination evidence absent
recorded unknown cannot establish
15high-signal profiles
12capability families
85primary sources
0Tracevity-tested system profiles
Local operational sequence
Inspect once. Gate every consequential change.
Tracevity Inspect runs locally, applies exact versioned evidence bindings, and returns concrete evidence, missing evidence, and proof ceilings. Tracevity Gate then compares baseline and candidate reconstruction outcomes to catch evidence loss across releases and migrations. Neither local command uploads traces or uses a model.
Compatibility intelligence
Follow what survives a directional trace-format mapping.
Tracevity separates documented interoperability, its own bounded interpretation, and provider-free fixture evidence. A generated destination artifact is not proof that a backend accepted it or that an external effect occurred.
The core question
We have traces. What can we actually reconstruct?
Trace volume is not evidentiary sufficiency. A detailed tool transcript can be excellent for debugging and still fail to identify the initiating principal, delegated authority, or resulting external state.
01Recorded
An event or artifact exists somewhere.
≠02Correlated
It can be joined to the right execution.
≠03Attributed
The actor or system identity is evidenced.
≠04Authorized
The governing authority is evidenced.
≠05Verified
The resulting external state is observed.
Direct answer
A useful reconstruction identifies the evidence available to the system, the observable decisions and actions that followed, and the state that resulted. It does not claim access to hidden chain-of-thought or true latent motivation.
Initial corpus
Read the trace ecosystem as evidence
Each profile separates documented capability, explicit gaps, Tracevity interpretation, and any separately linked Tracevity testing.
View all profilesOpenInference
An OpenTelemetry-based semantic convention for AI model, agent, retrieval, tool, graph, evaluation, and annotation traces.
2 documented families2 gaps or conflicts
W3C Trace Context
A W3C Recommendation defining interoperable HTTP trace-context propagation through traceparent and tracestate headers.
1 documented families9 gaps or conflicts
Model Context Protocol
A JSON-RPC protocol for connecting AI applications to tools, resources, prompts, and related capabilities, with W3C trace-context propagation.
0 documented families4 gaps or conflicts
OpenAI Agents SDK tracing
Built-in tracing for OpenAI Agents SDK workflows, including model calls, tools, handoffs, guardrails, and custom spans.
0 documented families3 gaps or conflicts
OpenTelemetry Protocol
The OpenTelemetry wire and file protocol for exporting trace, metric, and log telemetry between sources, collectors, and backends.
1 documented families8 gaps or conflicts
Reconstruction substrate
Twelve families. No single score.
“Good trace” only makes sense relative to a use case. Tracevity keeps each evidence family visible instead of compressing unlike questions into one number.
Identity
Which system and agent instance can be attributed?
Principal
Who initiated the activity, and under what delegated authority?
Context
What evidence and instructions were available to the system?
Decisions
Which explicit decision artifacts survived?
Models
What model interaction was recorded?
Tools
What was invoked, with what arguments, and what did the tool report?
Effects
Can a requested action be distinguished from a resulting external effect?
Human control
Which human interventions are represented?
Outcome
What happened, and who or what verified it?
Integrity
Can the evidence be altered or independently checked?
Portability
What meaning can move between tools, and what loss is documented?
Privacy
What sensitive material may persist, and which controls are documented?
Sufficient for what?
Start with the reconstruction job
01Basic debugging
Reconstruct the recorded model/tool sequence, arguments, responses, timing, and errors around a failed execution.
↗02Multi-agent reconstructionFollow work across parent agents, subagents, sessions, handoffs, and tool calls.
↗03Tool-action attributionAttribute a specific tool request and result to an agent execution and initiating principal.
↗04Human-delegation reconstructionEstablish which human or service delegated what authority, including approvals and elevation.
↗05External-effect verificationDistinguish requested operations and tool responses from provider acceptance and resulting destination state.
↗06Incident reconstructionBuild an evidence timeline across agent, identity, authorization, tool, and destination systems.
↗Operational artifact
Define the evidence you need before evaluating the trace you have.
Trace Reconstruction Requirements v0.2 adds capture-completeness evidence while the frozen v0.1 baseline remains downloadable.
Download the specification