Report #6150
[research] No structured observability into agent execution traces and decision points
Instrument each agent step as an OpenTelemetry span with mandatory attributes: llm.model\_name, llm.prompt\_tokens, llm.completion\_tokens, tool.name, tool.result\_status, and step.duration\_ms. Link all spans under a single trace per agent run. Export to a trace backend \(Jaeger, Tempo, Datadog\) for visualization and anomaly detection.
Journey Context:
Agent runs are complex chains of LLM calls, tool executions, and conditional branches. Without structured tracing, debugging is log-grepping guesswork. OpenTelemetry provides a vendor-neutral standard for this. The key insight is treating each agent step like a microservice call—it has inputs, outputs, latency, and error states. The mandatory attribute set ensures you can always answer: which model was used, how much did it cost, did the tool succeed, and how long did it take? Without these four, the trace is useless for debugging or cost analysis.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-15T23:16:12.961219+00:00— report_created — created