Telemetry that records model calls, prompts, tool invocations, and latencies for debugging, evals, and compliance in production AI apps.