Standard APM traces cannot distinguish a wrong tool choice from a hallucinated file path — the finalized OpenTelemetry GenAI semantic conventions define a standard attribute vocabulary (gen_ai.system, gen_ai.request.model, gen_ai.usage.input_tokens, gen_ai.tool.name) that Datadog, Langfuse, and Honeycomb all ingest natively. Emitting span events for high-cardinality data (user/assistant/tool messages) avoids blowing up attribute cardinality while preserving full fidelity. Derived metrics — context utilization (input tokens / context limit) alerting at 70%, tool error rate per tool name — give on-call engineers agent-specific signals alongside infrastructure metrics.