Observability

Tracing

• Every job gets a job_id and trace_id (OpenTelemetry)
• The Orchestrator emits a parent span for the whole interaction
• Worker jobs are child spans under the parent
• Tool cascade calls are sub-spans within the Orchestrator's span

A single user turn should be reconstructable end-to-end from one trace ID.

Logging

• Orchestrator: Tool cascade calls, decompose decisions, dispatch events, synthesis calls, re-plan events
• Workers: Job received, tool calls (with durations), completion, result summary
• All services: Structured JSON, indexed by session_id, job_id, trace_id

Metrics

Metric	Source	Use
Queue depth per agent type	Redis (LLEN)	Autoscaling signal
Job latency (enqueue → result)	Orchestrator job tracker	Primary SLO
Worker p50/p95/p99 latency	Worker self-report	Per-agent SLO
Tool cascade call count	Orchestrator	Cost and behavior tracking
Orchestrator synthesis token count	Orchestrator	Cost tracking
Schema validation failures	Orchestrator + worker	Quality signal

Alerts

• Queue depth sustained above threshold for a given agent
• Job p95 latency above SLO for a given agent
• Orchestrator re-plan rate spike (signal of brief quality or worker regressions)
• Redis connection errors
• Result schema validation failure rate above baseline

Cost Telemetry

The Orchestrator should log the per-turn token cost and tool call cost, broken down by phase (context load, dispatch, synthesis). This is the input to cost optimization. You can't optimize what you can't attribute.