What are the OpenTelemetry GenAI semantic conventions?
The OpenTelemetry GenAI semantic conventions are a shared set of span, event and attribute names, prefixed gen_ai, for recording model calls, agent runs, tool calls and eval results in a vendor-neutral telemetry format.
Checked
Every observability vendor once had its own way of describing a model call or a tool call, which made traces hard to move between tools. OpenTelemetry, the open standard most backend services already use for tracing, has a working group defining common names for AI telemetry so any compliant backend can read any compliant trace.
The conventions matter for agent teams because they decide whether a tracing setup can change vendor later without re-instrumenting the code.
What the conventions define
Agent operations
create_agent and invoke_agent spans for building and running an agent.
Tool calls
execute_tool spans for each tool the agent runs.
Retrieval and memory
retrieval, plus create_memory_store, search_memory and create_memory.
Eval results
A gen_ai.evaluation.result event with the evaluation name, a score value, a score label and an explanation.
Provider and protocol notes
Separate documents for OpenAI, Anthropic, AWS Bedrock and Azure, and for MCP and A2A.
Status as of October 2026
All GenAI conventions, including agent spans, events and attributes, are at Status: Development, not Stable. They moved into a dedicated repository, semantic-conventions-genai, created on 2026-05-05; the old page on opentelemetry.io now says the content has moved. The repository has no tagged releases yet and builds on core semantic conventions v1.44.0 (2026-08-04).
In practice this means attribute names can still change. Products that describe themselves as OTel-native are building on Development-status conventions, and instrumentation should be pinned to a known version.
How Motiontic's engineers apply it
Motiontic's engineers instrument agents with OpenTelemetry where the client's stack allows it, so traces can flow to the client's existing backend and to an AI-specific tool at the same time. Instrumentation versions are pinned and upgraded deliberately, with a check that dashboards and eval queries still read the renamed fields.
Main GenAI operations and what they record
| Operation | Records | Example |
|---|---|---|
| invoke_agent | One run of an agent | A support agent handling one ticket |
| execute_tool | One tool call and its result | Looking up an order by ID |
| retrieval | A search over a knowledge source | Fetching policy passages |
| search_memory | A read from long-term memory | Recalling a customer's past preferences |
| gen_ai.evaluation.result | An eval score attached to telemetry | A faithfulness score of 0.8 with an explanation |
Common questions
Are the OpenTelemetry GenAI semantic conventions stable?
No. As of October 2026 all GenAI conventions, agent spans included, are at Development status, and the dedicated repository has no tagged releases.
Where do the GenAI semantic conventions live now?
In the open-telemetry/semantic-conventions-genai repository on GitHub, created on 2026-05-05. The earlier pages on opentelemetry.io point to it.
Do the conventions cover evals?
Yes. The gen_ai.evaluation.result event carries an evaluation name, score value, score label and explanation, and Pydantic Logfire already emits it for its online evals.
Related
Agent observability
Step-by-step traces of what an agent did, why, and what it cost.
Trace grading
Scoring a whole recorded agent run, step by step.
Online evals
Grading live traffic in production, not only test sets before release.
Langfuse vs Arize Phoenix
Two self-hosted tracing and eval tools compared on licence, OpenTelemetry, evals and hosted options.
AI agent development
Agents that do real work for users, with the evals, traces and guardrails to keep them in production.
Running agents that have to hold up?
Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.