What is agent instrumentation?

Instrumentation is the code that records what an agent does while it runs and turns that record into telemetry your team can read later. For agents the closest thing to a standard is OpenTelemetry with the GenAI semantic conventions: an LLM call and a tool call each become a span, tagged with attributes like gen_ai.request.model or gen_ai.operation.name, so a reader, human or another system, can tell what happened without opening your source. A framework that already emits OTel spans exports them over OTLP, the wire protocol, to whatever endpoint is listening. Other dialects, OpenInference and Traceloop among them, describe the same events under different attribute names. Instrumentation is also where content decisions get made: what a span carries, and what gets stripped out before it leaves your process, is decided here, not downstream, once the data is already on the wire.