OpenTelemetry GenAI — Tracing Tool Calls End-to-End

The Problem

A debug from February 2026: user reports "my agent sometimes takes 30 seconds to respond; other times 3 seconds." No traces. Logs show the LLM call, but not the tool dispatch, not the MCP server round-trip, not the sub-agent. You guess. Eventually you find: one MCP server occasionally hangs on a cold-start.

Without end-to-end tracing, you cannot find this. OTel GenAI fixes it.

The conventions settled in 2025-2026 under the OpenTelemetry semantic-conventions group. They define stable attribute names so Datadog, Langfuse, Phoenix, OpenLLMetry, and AgentOps all parse the same spans. Instrument once; ship to any backend.

The Concept

Span hierarchy

Copyagent.invoke_agent  (top, INTERNAL span)
 ├── llm.chat       (CLIENT span)
 ├── tool.execute   (INTERNAL)
 │    └── mcp.call  (CLIENT span)
 ├── llm.chat       (CLIENT span)
 └── subagent.invoke (INTERNAL)

The whole thing nests under one trace id. Span ids link the parent-child relationships.

Required attributes

Per the 2025-2026 semconv:

• gen_ai.operation.name — "chat", "text_completion", "embeddings", "execute_tool", "invoke_agent".
• gen_ai.provider.name — "openai", "anthropic", "google", "azure_openai".
• gen_ai.request.model — requested model string (e.g. "gpt-4o-2024-08-06").
• gen_ai.response.model — the model actually served.
• gen_ai.usage.input_tokens / gen_ai.usage.output_tokens.
• gen_ai.response.id — provider response id for correlation.

For tool spans:

• gen_ai.tool.name — tool identifier.
• gen_ai.tool.call.id — the specific call id.
• gen_ai.tool.description — tool description (optional).

For agent spans:

• gen_ai.agent.name / gen_ai.agent.id / gen_ai.agent.description.

Span kinds

• SpanKind.CLIENT for calls crossing a process boundary (LLM provider, MCP server).
• SpanKind.INTERNAL for the agent's own loop steps and tool execution.

Opt-in content capture

By default, spans carry metrics and timing — not prompts or completions. Large payloads and PII are off by default. Set OTEL_SEMCONV_STABILITY_OPT_IN=gen_ai_latest_experimental and specific content-capture env vars to include content. Review carefully before enabling in prod.

Events on spans

Token-level events can be added as span events:

• gen_ai.content.prompt — input messages.
• gen_ai.content.completion — output messages.
• gen_ai.content.tool_call — tool call as recorded.

Events time-order within a span for detailed replay.

Exporters

OTel spans export to:

• Jaeger / Tempo. OSS, on-prem.
• Langfuse. LLM-observability-specific; visualizes token usage.
• Arize Phoenix. Evals + tracing combined.
• Datadog. Commercial; natively parses gen_ai.* attributes.
• Honeycomb. Column-oriented; query-friendly.

All speak OTLP, the wire format. Your code does not care.

Propagation across MCP

When an MCP client calls a server, inject the W3C traceparent header into the request. Streamable HTTP supports standard headers. Stdio does not carry HTTP headers natively; the spec's 2026 roadmap discusses adding a _meta.traceparent field on JSON-RPC calls.

Until that ships: include the traceparent in the _meta of every request manually. Server logs the trace id.

Metrics

Alongside spans, the GenAI semconv defines metrics:

• gen_ai.client.token.usage — histogram.
• gen_ai.client.operation.duration — histogram.
• gen_ai.tool.execution.duration — histogram.

Use these for dashboards that do not need per-call detail.

AgentOps layer

AgentOps (founded 2024) specializes in GenAI observability. It wraps popular frameworks (LangGraph, Pydantic AI, CrewAI) to emit OTel spans automatically. Useful if your stack uses a supported framework; use manual instrumentation otherwise.

Use It

code/main.py emits OTel-shaped spans to stdout (in OTLP-JSON-like format) for an agent that calls an LLM, dispatches two tools, and makes one MCP round-trip. No real exporter — the lesson focuses on the span shape and attribute set. Paste the output into an OTLP-compatible viewer or just read it.

What to look at:

• Trace id is shared across all spans.
• Parent-child links are encoded via parentSpanId.
• Required gen_ai.* attributes are populated.
• Content capture is off by default; one scenario turns it on via env var.

Ship It

This lesson produces outputs/skill-otel-genai-instrumentation.md. Given an agent codebase, the skill produces an instrumentation plan: where to add spans, which attributes to populate, and which exporters to target.

Exercises

1. Run code/main.py. Count the spans and identify which is CLIENT vs INTERNAL.

2. Turn on content capture (env var) and confirm gen_ai.content.prompt and gen_ai.content.completion events appear. Note the implications for PII.

3. Add the tool-execution metric gen_ai.tool.execution.duration and emit it as a histogram sample per call.

4. Propagate a traceparent from a parent agent span into an MCP request's _meta.traceparent field. Verify the MCP server would see the same trace id.

5. Read the OTel GenAI semconv spec. Identify one attribute listed in the semconv that this lesson's code does NOT emit. Add it.

Key Terms

Further Reading

• OpenTelemetry — GenAI semconv — canonical conventions for GenAI spans, metrics, and events
• OpenTelemetry — GenAI spans — LLM and tool-execution span attribute list
• OpenTelemetry — GenAI agent spans — agent-level invoke_agent span
• open-telemetry/semantic-conventions — GenAI spans — GitHub-hosted source of truth
• Datadog — LLM OTel semantic convention — production integration walkthrough

Diagrams1 figure