Design tracing for a multi-step agent where the failures are silent. The run returns a plausible answer that is wrong.

This separates people who have debugged agents in production from people who have read about tracing. The naive answer logs the prompt and the final response and calls it observability; it cannot tell you that step 3's retrieval returned zero rows and the model confabulated around the gap. The discriminating instinct is that the dangerous failures are the ones that do not raise an exception: empty tool results, truncated context, a guardrail that rewrote the output, plan drift where the agent picks the wrong branch. The follow-up that scores: 'a user reports a wrong answer, walk the trace.' A strong candidate names span attributes (tool args, raw output, retry count, token usage, finish_reason) and the outcome signal (eval score, user thumbs, downstream correction) that lets them find these runs at scale instead of one ticket at a time. Saying 'we use OpenTelemetry' without naming what to put on the spans is a tell.