observability · Workflow as an Agent

What this lesson demonstrates

Two hosted agents — “French” then “English” — are chained by a workflow graph (French → English). Telemetry is switched on at build time via WithTelemetry, so every executor invocation and edge traversal opens an OpenTelemetry span. The finished workflow is then wrapped by agentworkflow.NewAgent, which turns the whole graph into one *agent.Agent: you call RunText on it exactly like a plain agent, and the two-stage pipeline runs underneath.

The real code

Telemetry is a builder step, right alongside the edges and outputs:

wf, err = workflow.NewBuilder(frenchBinding).
    AddEdge(frenchBinding, englishBinding).
    WithOutputFrom(englishBinding).
    WithTelemetry(
        workflowotel.New(workflowotel.Config{SourceName: sourceName}),
        workflow.TelemetryOptions{EnableSensitiveData: true},
    ).
    Build()

Then the whole workflow becomes one agent:

return agentworkflow.NewAgent(wf, agentworkflow.AgentConfig{IncludeOutputsInResponse: true})

Callers run it with a plain wfAgent.RunText(ctx, "...", agent.WithSession(session), agent.Stream(true)) — the French→English pipeline executes invisibly.

What to notice

• SourceName is the instrumentation scope. "Workflow.Sample.WorkflowAsAgent" is the name a trace backend groups these spans under — the equivalent of a tracer name.
• EnableSensitiveData: true is a deliberate choice. It lets prompt/response content ride along in spans. Useful in a demo, but a real deployment weighs that against not logging user data.
• agentworkflow.New vs. agentworkflow.NewAgent. The first binds a single agent as a workflow executor (a node); the second wraps a whole workflow as one agent. This lesson uses both — two nodes built with New, the graph wrapped with NewAgent.
• IncludeOutputsInResponse surfaces the final output. Without it, the wrapped agent runs the pipeline but the terminal English output wouldn’t appear in the response stream.
• Port note: the OpenTelemetry stdout exporter isn’t in this repo’s module graph, so the lesson keeps the concept — workflowotel.New + WithTelemetry — but relies on the global no-op tracer instead of exporting spans, and uses a tiny local logging middleware so each turn is still visible.

How it maps to the Agent Framework

In the Microsoft Agent Framework Go SDK, workflow/observability/opentelemetry is the standard way to get distributed traces across a multi-agent graph — spans nest so you can see the French turn, the English turn, and the edge between them in a backend like Azure Monitor. Wrapping a workflow as an agent is the composition trick that lets a whole pipeline plug into anything expecting an agent.Agent. Running it live needs az login, FOUNDRY_PROJECT_ENDPOINT, and FOUNDRY_MODEL because both stages call the Azure AI Foundry model.

Run it

go run ./tutorial/03-workflows/observability/workflow_as_an_agent

Runs live (both agents call Foundry); the offline test builds the same graph with a fake credential and asserts the French→English wiring with no network.

Next: shared-states · Coordinating executors through shared state