OpenTelemetry GenAI Export

AgentSight can export the LLM calls it captures as OpenTelemetry GenAI semantic-convention (gen_ai.*) spans, sent to any OpenTelemetry Collector over OTLP/HTTP. Because AgentSight reconstructs the traffic from the kernel (eBPF) side, you get standards-compliant GenAI telemetry for any agent — including closed-source CLIs — with zero in-process instrumentation.

Spec: https://opentelemetry.io/docs/specs/semconv/gen-ai/

Quick start

Point AgentSight at a running collector with --otel:

sudo ./agentsight debug trace --otel --otel-endpoint http://localhost:4318

Each LLM request/response pair becomes a chat {model} CLIENT span. By default only metadata is exported; pass --otel-capture-content to also include the prompt/completion text (off by default for privacy).

Flags

Standard OTel env vars are honored: OTEL_EXPORTER_OTLP_ENDPOINT, OTEL_EXPORTER_OTLP_TRACES_ENDPOINT (used as-is if set), and OTEL_SERVICE_NAME (default agentsight).

Run a collector

A minimal collector that prints received spans:

# otel-collector-config.yaml
receivers:
  otlp:
    protocols:
      http:
        endpoint: 0.0.0.0:4318
exporters:
  debug:
    verbosity: detailed
service:
  pipelines:
    traces:
      receivers: [otlp]
      exporters: [debug]

docker run -d --name otelcol -p 4318:4318 \
  -v $(pwd)/otel-collector-config.yaml:/etc/otelcol/config.yaml \
  otel/opentelemetry-collector:latest

Integrating with other tools

Because the export is plain OTLP, anything in the OpenTelemetry ecosystem can be the backend — just change the collector's exporters:

• Jaeger — Jaeger accepts OTLP natively (ports 4317/4318); point AgentSight at Jaeger, or fan out from the collector with an otlp exporter.
• Grafana Tempo — add an otlp exporter targeting Tempo.
• Vendors (Datadog, Honeycomb, New Relic, …) — all consume the OTel GenAI conventions; add the vendor exporter to the collector pipeline.

Example collector pipeline forwarding to a vendor:

exporters:
  otlphttp/vendor:
    endpoint: https://otlp.example-vendor.com
    headers: { "api-key": "${VENDOR_API_KEY}" }
service:
  pipelines:
    traces:
      receivers: [otlp]
      exporters: [otlphttp/vendor]

Emitted attributes

Per the GenAI agent/model span conventions:

The span's start/end times come from the captured request and response timestamps, so latency is measured at the wire.

How it works

agent → SSL_write/SSL_read (captured by eBPF sslsniff)
      → HTTPParser reconstructs request + response
      → OtelExporter pairs them by (pid, tid), maps to gen_ai.* attributes
      → POST OTLP/HTTP JSON to the collector's /v1/traces

Provider detection is host-based and falls back to the host name for OpenAI-compatible endpoints (self-hosted vLLM, llama.cpp, …), so those still produce useful spans. Streamed (SSE) responses whose body can't be reparsed as JSON still emit a span with the request attributes and HTTP status.

Notes & limitations

• OTLP/HTTP only (the standard collector receiver). For a remote collector behind TLS, run a local collector and let it forward upstream.
• One span per request/response pair (chat); multi-agent workflow spans (invoke_agent, execute_tool) are not yet emitted — see issue #47.
• The GenAI conventions are still experimental; attribute names track the current spec.