EmissarRequest early access
Menu

Glossary · Agent architecture

OpenTelemetry

The CNCF's vendor-neutral standard and toolkit for traces, metrics and logs. Its GenAI and agent semantic conventions are still in Development status.

By Emissar team · Published · Updated

OpenTelemetry (OTel) is an open-source, vendor-neutral observability framework, made up of a specification, APIs, SDKs and tools, for generating, collecting and exporting telemetry such as traces, metrics and logs.

Components. The project defines a specification, the OTLP protocol for sending telemetry, semantic conventions that fix attribute names, language SDKs and instrumentation libraries, and the Collector, which receives, processes and forwards data. OpenTelemetry is not a storage or analysis backend. It sends data to tools such as Jaeger, Prometheus or commercial services.

Status. OpenTelemetry formed in 2019 from the merger of OpenTracing and OpenCensus. The Cloud Native Computing Foundation accepted it on May 7, 2019, moved it to incubating on August 26, 2021, and graduated it on May 11, 2026. The core semantic conventions are released regularly; the latest release on GitHub at the time of writing is v1.44.0, from August 4, 2026.

GenAI semantic conventions. Conventions for generative AI now live in their own repository, semantic-conventions-genai, and the core repository marks its old GenAI pages as moved. Their status, as of September 26, 2026, is Development, so names can still change. They cover model spans, agent spans, events, exceptions, metrics and MCP, plus pages for specific providers: Anthropic, AWS Bedrock, Azure AI Inference and OpenAI.

  • Agent spans set gen_ai.operation.name to values such as create_agent, invoke_agent, invoke_workflow or execute_tool.
  • Common attributes include gen_ai.provider.name, gen_ai.agent.name, gen_ai.agent.id and gen_ai.conversation.id.
  • Message content (gen_ai.input.messages, gen_ai.output.messages, gen_ai.system_instructions) is opt-in, and the conventions warn it is likely to contain sensitive information.

Why it matters for agents. A shared trace format lets one run be followed across frameworks and vendors. Google ADK emits OTLP traces that follow the GenAI agent conventions, and Microsoft Agent Framework documents OpenTelemetry instrumentation based on them. Because the conventions are still in Development, pin the version your instrumentation targets.

Neighbouring terms. Tracing is the practice; OpenTelemetry is the most common format for it. Agent evaluation often reads OpenTelemetry traces.

Sources

  1. OpenTelemetry documentation: What is OpenTelemetry? (accessed )
  2. CNCF: OpenTelemetry project page (maturity level and dates) (accessed )
  3. OpenTelemetry semantic conventions releases on GitHub (accessed )
  4. OpenTelemetry GenAI semantic conventions repository (accessed )
  5. OpenTelemetry GenAI semantic conventions: Agent spans (accessed )
  6. OpenTelemetry GenAI semantic conventions: Model spans (accessed )
  7. Agent Development Kit documentation: Traces (accessed )
  8. Microsoft Learn: Agent observability (Microsoft Agent Framework) (accessed )