PARAS MADAN
All writing

Latest in AI

Azure Monitor Cross-Agent Tracing: Find the Missing Handoff

A practical debugging checklist for Azure Monitor's connected agent traces, including instrumentation, context propagation, content capture and preview limits.

Microsoft announced connected cross-agent tracing in Azure Monitor on November 18. When an agent returns an incomplete answer, a clean-looking dashboard can still be misleading. The failing tool might be outside the portion of the workflow you are collecting.

Azure Monitor's November 18 cross-agent tracing update makes connected investigations more useful. It does not remove the need to check which components emitted telemetry and whether the handoff between them stayed connected. I would verify that evidence before changing a model or rewriting a prompt.

Know which parts can participate

Microsoft's announcement describes Application Insights Agents as preview. Copilot Studio, Foundry and GitHub Copilot agents are supported; external agents need compatible OpenTelemetry GenAI telemetry and trace-context propagation. Microsoft 365 Copilot and WorkIQ use the preview A365 exporter. Fabric agent tracing is listed for December 2026.

A trace covers participating components and the telemetry they actually emit. Tool inputs and outputs appear only where those details are emitted and collection is enabled. Read the tracing announcement.

That is the boundary I would write into the rollout plan: connected visibility is conditional, not automatic coverage of every agent involved.

Draw the expected path before opening the trace

My first debugging step would be to sketch one real request: entry point, first agent, delegated agent, external tool and response. Beside each step, record its owner and telemetry destination.

This gives you an expected execution path to compare with the observed one. Without it, a trace that ends early can look like a successful short run rather than a missing service.

Choose a request whose expected result is easy to verify. Use non-sensitive test data, and retain a timestamp and request identifier so the application owner and observability owner investigate the same run.

Distinguish missing instrumentation from broken correlation

Microsoft's client-side tracing guide explains that Foundry's portal-side traces do not replace instrumentation in your own application. Client-side OpenTelemetry can capture model calls, tool invocations and custom logic. GenAI tracing requires an explicit opt-in.

I would diagnose two different failures. If a service emits no relevant spans, inspect its instrumentation and export configuration. If it emits spans that appear as a separate run, inspect context propagation at the boundary.

A useful handoff test is to follow the same request through both sides of a delegation. Do not conclude the downstream service was never called merely because it is absent from the first view.

Keep payload capture a separate decision

The same guide treats content recording separately from enabling tracing. It also marks the client-side feature as experimental preview, without an SLA and not recommended for production workloads. SDK behavior and emitted attributes can change. Review the setup and preview caveats.

My default investigation would start with operation names, durations, status and correlation. Turn on message or tool-content capture only after deciding what may be recorded, who may read it and how long it should remain.

More detail is not always better evidence. A trace full of sensitive payloads creates a second problem if a timing signal would have answered the original question.

Validate a failure, not just a happy path

For a controlled pilot, I would exercise a successful request and an intentional tool failure in a non-production environment. Check whether the failure appears at the correct dependency and whether the parent run shows the resulting outcome.

Then test a delayed response and a retry. The question is whether an engineer can distinguish waiting, repeated attempts and model processing rather than blaming every slow response on inference.

These are proposed checks, not tests I have performed against the preview.

The acceptance criterion should be practical: can the service owner locate the failing handoff from the evidence available? If not, fix the telemetry gap before using the dashboard to compare models, estimate savings or declare the agent reliable.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help AI infrastructure teams explain observability and failure handling through practical developer stories.

Book a 30-minute call.