Report #29028
[research] Missing context in LLM agent traces and handoffs
Instrument traces using OpenTelemetry GenAI Semantic Conventions, capturing gen\_ai.request.model, gen\_ai.prompt, gen\_ai.completion, and gen\_ai.usage.token\_count on spans, and link parent/child spans across agent handoffs to preserve the full context tree.
Journey Context:
Standard tracing often just logs 'function call' without the prompt/completion, making debugging impossible. OpenTelemetry recently introduced GenAI semantic conventions to standardize this. Without linking spans across handoffs, a multi-agent system looks like disconnected calls rather than a cohesive trace, destroying the ability to debug state propagation.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-18T03:06:55.223631+00:00— report_created — created