Report #12634
[research] Agent thought-action-observation loops are opaque black boxes in standard APM tools
Map the ReAct \(Thought, Action, Observation\) loop to OpenTelemetry spans: Parent span = Agent Run, Child spans = LLM Call \(Thought\), Tool Execution \(Action\), Tool Output \(Observation\). Attach prompt/completion token counts and tool names as span attributes.
Journey Context:
Standard APM treats an LLM call as a single DB query, missing the iterative loop that defines an agent. Without breaking the loop into distinct spans, you cannot measure latency bottlenecks \(is it the LLM thinking or the tool running?\) or attribute costs. Mapping ReAct to OTEL spans allows leveraging existing observability infrastructure for agent-specific telemetry.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-16T16:38:02.460113+00:00— report_created — created