Observability

Prometheus Metrics

Metric	Type	Labels
agentops_requests_total	Counter	route, status
agentops_request_latency_seconds	Histogram	route
agentops_tool_execution_total	Counter	tool_name, status
agentops_tool_execution_latency_seconds	Histogram	tool_name
agentops_verifier_escalations_total	Counter	—

Structured Logging

Every log line is a flat JSON object with trace_id injected automatically via contextvars.ContextVar:

{
  "timestamp": "2026-07-03T10:00:00.000Z",
  "level": "info",
  "event": "request_received",
  "trace_id": "550e8400-e29b-41d4-...",
  "customer_id": "cust_123"
}

Tracing

Set OTEL_ENABLED=true and OTEL_ENDPOINT to export per-stage spans (classify/plan/execute/verify) to any OTLP-compatible backend (Jaeger, Tempo, Honeycomb, etc.). app/telemetry wraps OpenTelemetry behind a no-op shim, so every start_as_current_span() call is a zero-cost context manager when tracing is disabled.

Execution Traces

Every /messages request produces a structured Trace in TraceStore: state transitions, tool calls, and agent decisions, retrievable via GET /trace/{trace_id}.