How do I trace a Claude Agent SDK agent?

Set CLAUDE_CODE_ENABLE_TELEMETRY=1 and an OTLP exporter for each signal you want. The SDK runs the Claude Code CLI as a child process, and the CLI does the actual exporting: metrics (OTEL_METRICS_EXPORTER) for token, cost, and session counters, structured log events (OTEL_LOGS_EXPORTER) for each prompt, API call, and tool result, and traces (OTEL_TRACES_EXPORTER) for spans around each interaction, model request, tool call, and hook. Traces are the one signal still in beta, so they also need CLAUDE_CODE_ENHANCED_TELEMETRY_BETA=1.

A subagent’s spans nest under the parent’s tool span, so a delegated task shows up inside the same trace as the loop that spawned it, rather than as a disconnected run. What doesn’t come free is the content: prompt text, tool arguments, and full request bodies are opt-in, one environment variable each, so a fresh setup gives you shape and timing before it gives you what the agent actually said. Grading what it said is a separate job, closer to replaying the turn that failed with its real context than to reading a span.