Why production AI agents need tracing: spans, not just outputs

A support agent gave a customer a wrong shipping date, and the team could not tell from the transcript whether the cause was the model, a stale lookup, or the prompt. The article argues that agents in production need the same kind of tracing that distributed software adopted years ago. Each run would be recorded as a nested tree of timed spans covering model calls, tool calls, retrievals and guardrail checks, so failures can be pinpointed to a single step. The article notes that OpenTelemetry's GenAI conventions are standardising these span names and attributes across frameworks and providers. It also separates cheap, safe metadata such as durations and token counts from sensitive content such as prompts and tool outputs, which carries the debugging value but also customer data risk.