Skip to content
Olivares AI docs

OpenTelemetry GenAI (any instrumented agent)

Claude Code is the canonical cooperative source, but it is not the only cooperative agent you run. The same connector that receives Claude Code’s telemetry (kind: claude) carries an opt-in, vendor-neutral OpenTelemetry GenAI profile: point any OTel-instrumented agent or framework at the same OTLP receiver, and its gen_ai.* telemetry feeds the access-map and cost pipeline — LangChain, LangGraph, CrewAI, AutoGen, Google ADK and anything else that emits the GenAI semantic conventions on spans or log events.

Why it is opt-in

Section titled “Why it is opt-in”

The OpenTelemetry GenAI conventions are Development status (pre-stable) upstream, and three dialects genuinely coexist in 2026 fleets. So the profile is off by default and gated exactly like the OTel SDKs gate it — by the opt-in token:

{
  "sources": [{
    "name": "agents-otel",
    "kind": "claude",
    "tenant": "<tenant-id>",
    "config": {
      "semconv_opt_in": "gen_ai_latest_experimental"
    }
  }]
}

semconv_opt_in mirrors OTEL_SEMCONV_STABILITY_OPT_IN: a comma-separated list that must contain gen_ai_latest_experimental. With the profile off, a gen_ai.* record still feeds the session-liveness watchdog but is not mapped — honest absence, not silent ingestion.

What the normalizer accepts

Section titled “What the normalizer accepts”

The profile is pinned to semconv v1.41.1 and normalizes the three GenAI dialects that coexist in real estates, stamping every normalized event with the dialect’s semconv pin so provenance survives:

DialectShape
Legacy OpenLLMetryindexed gen_ai.prompt.{i}.* attributes
v1.36 and priorthe deprecated per-message events
v1.37+the messages generation

On top of the message shapes it maps the mcp.* conventions (v1.39) and the invoke_agent client/internal split plus invoke_workflow (v1.41) — so framework-orchestrated agent and workflow invocations land as structured topology, not noise. Span-based emission (how LangGraph, LangChain, CrewAI, AutoGen and Google ADK instrument) and log-based emission are both ingested.

Cost samples are de-duplicated by W3C span id, so an agent whose telemetry arrives on both the span and log paths is never double-billed.

Wiring an agent to it

Section titled “Wiring an agent to it”

The receiver is the connector’s own OTLP endpoint (gRPC 127.0.0.1:4317, HTTP 127.0.0.1:4318 by default). On the agent side, standard OTel SDK configuration applies — exporter endpoint to the loopback receiver, and the GenAI opt-in if your instrumentation gates on it:

Terminal window
OTEL_EXPORTER_OTLP_ENDPOINT=http://127.0.0.1:4318
OTEL_SEMCONV_STABILITY_OPT_IN=gen_ai_latest_experimental

What you’ll see in the console

Section titled “What you’ll see in the console”

Instrumented sessions appear in Sessions as live activity, attributed to the emitting agent; their model calls feed Cost & FinOps; MCP and tool spans contribute edges to the Access map like any cooperative source:

The Sessions view showing live agent session activity from cooperative telemetry. The Sessions view showing live agent session activity from cooperative telemetry.

Honest limits

Section titled “Honest limits”
  • Pre-stable conventions, pinned ingest. The profile is pinned to v1.41.1; when upstream moves, the pin moves by a deliberate update, not by silent drift. Instrumentation that emits a fourth dialect is not guessed at.
  • Cooperative means cooperative. An agent that does not emit is invisible to this path — that is what eBPF/Tetragon and store-native audit are for.
  • Framework span-kind quirks are real. Some frameworks emit spans whose kind does not match the v1.41 client/internal rules; the normalizer maps what it can prove and leaves the rest unmapped rather than misattributed.
Section titled “Related”