OpenTelemetry GenAI Semantic Conventions
OpenTelemetry GenAI Semantic Conventions are shared telemetry definitions for generative-AI systems. The official repository extends core OpenTelemetry semantic conventions with spans, metrics, and events for GenAI clients, agents, Model Context Protocol activity, and provider-specific integrations. The conventions standardize names and structures; they are not a monitoring backend, evaluation method, or guarantee that an instrumented value is complete or correct.
Origin and context
OpenTelemetry described its GenAI conventions and instrumentation work in December 2024. Datadog announced native ingestion of GenAI spans in December 2025. By May 2026, Microsoft documented Agent Framework tracing into Foundry, and an OpenTelemetry walkthrough demonstrated the conventions through VS Code Copilot and Aspire. The separate GenAI repository contains specification models, generated documentation and reference scenarios. This is cross-ecosystem implementation evidence, not proof that all signals or conventions have reached stable status.
Why it matters
A shared field vocabulary lets a model call remain recognizable when telemetry passes from an application's instrumentation to a collector and a backend. Engineers can relate an agent invocation to child model and tool operations and inspect timing or token counts without treating each provider's naming as a separate language. Portability is conditional on compatible convention versions and emitted fields. The schema supplies meaning for recorded data; it does not decide whether an answer is correct or whether all relevant work was instrumented.
Example
An illustrative agent run contains a parent invocation, a model call and a search-tool call. Compatible instrumentation records related spans with the operation, model and token-usage attributes. A backend can show where the time was spent and how the calls relate. Capturing the messages themselves is a separate choice: the OpenTelemetry May 2026 walkthrough keeps content capture distinct from metadata. An unstructured prompt log alone cannot provide the same operation hierarchy and shared field semantics.
How it differs
Agent observability
Agent observability is the broader practice of understanding an agent's behavior, quality, cost, and failures through traces, logs, metrics, evaluations, and operational context. OpenTelemetry GenAI Semantic Conventions are one schema-level building block for that practice. Conforming spans improve interoperability, but they do not choose evaluation criteria, detect hallucinations, or explain a failed decision without additional instrumentation and analysis.
Maturity and evidence
Maturity remains 3. Datadog ingestion, Microsoft Agent Framework tracing and OpenTelemetry's reference tooling establish a concrete, shared convention family used outside its originating repository. The lifecycle is established for that identifiable schema family, not a declaration that its specification is stable. Some documented integrations are previews, coverage varies and versions can differ. Those limitations prevent assuming universal conformance or raising the rating on the strength of vendor announcements alone.
Limits and open questions
Prompt and tool content can contain sensitive information and is not synonymous with basic latency or token telemetry. The OpenTelemetry walkthrough makes content capture opt-in; that implementation detail should not be assumed for every instrumentor. Backends can support different convention versions, and experimental signals can change. Standardized names also cannot repair inaccurate values, missing spans or incompatible instrumentation, so schema compliance and application-quality evaluation remain separate checks.
Related terms
References
- OpenTelemetry for Generative AIOpenTelemetry · 2024-12-05 · class A
- Datadog Agent Observability natively supports OpenTelemetry GenAI Semantic ConventionsDatadog · 2025-12-01 · class B
- OpenTelemetry GenAI Semantic Conventions README (version 2026-06-04)OpenTelemetry · 2026-06-04 · class A
- Inside the LLM Call: GenAI Observability with OpenTelemetryOpenTelemetry / Microsoft · 2026-05-14 · class A
- What's new in Microsoft Foundry | April 2026Microsoft · 2026-05-12 · class A
Last updated: 2026-09-05
This term is also covered in the Skills Atlas as opentelemetry skill.
This term is also covered in the Skills Atlas as llm observability skill.