OpenTelemetry gen_ai.* Semantic Conventions

OpenTelemetry (OTel) is the CNCF-graduated observability standard: a vendor-neutral API, SDK, and protocol (OTLP) for distributed tracing, metrics, and logs. The gen_ai.* semantic conventions are a set of standardized attribute names and span schemas for observing AI/LLM workloads — the extension of OTel into the agentic-AI observability space.

What the gen_ai.* conventions define

gen_ai.* SemConv (v1.37+ as of May 2026; status: experimental but broadly adopted) specifies:

• Span names — gen_ai.client.request, gen_ai.client.response, gen_ai.tool.call, gen_ai.retrieval
• Standard attributes — gen_ai.system, gen_ai.request.model, gen_ai.usage.input_tokens, gen_ai.usage.output_tokens, gen_ai.tool.name, gen_ai.tool.call.id
• Agent-specific spans — gen_ai.agent.step, gen_ai.agent.invocation for multi-step agent traces

SIG contributors as of May 2026: Amazon, Elastic, Google, IBM, Langtrace, Microsoft, OpenLIT, Scorecard, Traceloop. The multi-stakeholder SIG is the primary signal of standard status — no single vendor can capture it.

Why OTel is the foundational observability choice

• Vendor-neutral — OTel traces can be sent to any backend (Datadog, Grafana, Splunk, Jaeger, Honeycomb, etc.) without changing instrumentation code. Lock-in is at the backend level, not the collection level.
• Already in the stack — most organizations already instrument microservices with OTel; adding gen_ai.* spans extends existing infrastructure rather than creating a parallel observability silo.
• No license cost — the OTel SDKs (Python, JS, Go, Java, etc.) are Apache 2.0. The gen_ai.* SemConv is a specification, not software — adopting it has zero cost.
• CNCF graduation — OTel is a CNCF Graduated project, the highest maturity level. It has production adoption at Google, Microsoft, AWS, Meta, Netflix, and others.
• Agent-specific volume — agents generate 10–20× the log volume of human users over the same window. OTel’s pre-aggregation-at-hook model (span batching, tail sampling) is the right architecture for this volume.

Agent observability use pattern

In an agentic-AI system, OTel gen_ai.* spans flow through the six planes of the RA:

Agent process
  ├── gen_ai.client.request span → model call
  ├── gen_ai.tool.call span → tool invocation (→ Egress plane)
  ├── gen_ai.retrieval span → RAG retrieval (→ Data plane)
  └── gen_ai.agent.step span → per-step trace (→ Observability plane)

Each span carries agent_id, user_id, session_id attributes (via the Agent Observability §3 identity-multiplexing pattern), making every action traceable to a human principal.

In the RA / CMM

• RA Observability Plane: OTel gen_ai.* SemConv is the primary reference implementation for the Observability Plane — classified as Std (CNCF standard).
• CMM D7 L3: “OTel gen_ai.* semantic conventions emitted across agents” is the minimum evidence artifact for Defined observability.
• CMM D9 L3: OTel latency/cost spans are used to measure guardrail latency (p50/p95/p99) per agent.
• FOSS/small-team stack: OTel is the recommended zero-cost observability foundation; backend can be Langtrace/Traceloop (OSS) or any OTel-compatible SaaS.
• Enterprise stack: OTel spans feed into existing SIEM (Splunk, Datadog, Dynatrace, etc.) without replatforming.

Implementations / backends

Tool	Type	Role
Langtrace	OSS	Agent-aware OTel tracing with LLM-specific UI
Traceloop	OSS	OpenLLMetry SDK (OTel-native)
Helicone	OSS / SaaS	LLM observability gateway; OTel-compatible
LangSmith	SaaS	LangChain-native; exports OTel
DataDog AI Monitoring	SaaS	OTel-native ingestion; AI-specific dashboards
New Relic AI Monitoring	SaaS	OTel-native; gen_ai.* support

See also

• Agent Observability — the wiki’s observability practice page, which uses OTel as its foundation
• Agentic AI Security RA §Observability plane
• Agentic AI Security CMM 2026 D7 + D9