Instrumentation

Instrumentation is the code that records what an agent does while it runs and emits that record as telemetry. For agents, the standard is OpenTelemetry with the GenAI semantic conventions. A framework that already emits OTel spans exports them over OTLP, via HTTP or gRPC, to any endpoint that accepts the protocol. Other dialects, such as OpenInference and Traceloop, describe the same events with different attribute names.

Agent telemetry has a three-level structure that follows the shape of agent execution. A session groups everything that happened across one conversation. A trace covers one turn: a user message, the agent's full response, and everything between. A span covers one step within the turn, an LLM call or a tool call. Spans nest within traces, and traces group into sessions.

Instrumentation is also the point where content leaves the application. It decides what gets emitted: which prompts, outputs, and tool arguments appear in the record, and in what form. The telemetry downstream is whatever instrumentation chose to emit, so content decisions, including redaction, happen here.