Agent Observability and Telemetry (OpenTelemetry GenAI)

An agent run is a tree of model calls, tool calls, retrievals and memory reads, and observing it means tracing that tree, measuring it at the call, run and fleet level, and attaching the outcome. OpenTelemetry's GenAI conventions now name the spans and metrics (invoke_agent, execute_tool, tool calls per run), which makes traces portable across vendors, but they are still marked Development and content capture is opt-in for good reason.