Trace multi-agent workflows with OpenTelemetry

Why it matters

When agents cross service boundaries, one missing context injection splits the trace, and operators lose the waterfall that names the responsible agent.

What to practise

Create OpenTelemetry spans for a synthetic router, recommendation, and inventory flow, and inject and extract the W3C traceparent at each handoff. Confirm three spans share one trace ID, then export spans and structured logs to Application Insights with Microsoft Entra authentication. Inspect the Agents (Preview) view, the transaction waterfall, and the supplied KQL queries. Move latency, error-rate, and token thresholds in YAML and watch the anomaly decisions change without editing KQL. Finish by watching a scheduled-query alert fire, notify its action group, and auto-resolve. No model is called; the flow uses synthetic metadata.

How it fits

Trace-context propagation and privacy-aware decision logs are 3.3.2. Per-agent latency, errors, and chain error rate are the health measures of 3.3.1, and the alert that consumes emitted policy decisions is 3.3.4.