Agentic systems are still just code. The requests move through your system, so the distributed tracing you already use still applies. What's different is the shape of the telemetry and the non-determinism: one user turn fans out into a loop of model calls, tool calls, retrievals, and retries that branch on whatever the model just decided. To understand what an agent is doing, and to answer for its cost, latency, and quality, you have to see the whole loop, not any single call.

So you go looking, and the telemetry isn't there. The agent picked a tool, retried twice, burned tokens, and returned a plausible answer. You see a 200 OK. Tool calls fail quietly: the model gets a bad result, keeps going, and the failure surfaces as a strange answer, not an error span. Auto-instrumentation captures the HTTP call to the model and stops, so the reasoning, the retries, and the tool decisions that ran up the cost stay invisible.

This talk surveys the OpenTelemetry GenAI semantic conventions (the gen_ai.* attributes, a turn modeled as a trace, a conversation that links turns) and the best practices for capturing them, so you stop guessing at what your agents are doing. The conventions are still evolving, we'll touch on strategies for instrumenting what they don't yet.
Sobre Wolfgang Therrien: Wolfgang is a Staff Engineer and Technical Lead for Frontend Observability at Honeycomb.io and a maintainer of the OpenTelemetry browser repository. Show less