workflow_as_an_agent · A Workflow, Wrapped as One Agent

What this lesson demonstrates

Two Foundry agents — French and English — are wired into a concurrent workflow that fans one prompt out to both in parallel. Then agentworkflow.NewAgent hosts that whole workflow as a single agent (BilingualWorkflowAgent). The caller in main never touches the graph: it calls RunText on one agent and receives the merged bilingual answer. This is the composition payoff — any workflow you can build can be collapsed into an agent, and then dropped inside a larger workflow, because it now is an agent.

Wrapping the workflow

Construction is factored into buildWorkflowAgent(...) so the offline test can build the identical structure with a fake credential. The concurrent workflow is built first, then hosted:

wf, err := agentworkflow.NewConcurrentWorkflowBuilder(french, english).
    WithName("bilingual-workflow").
    Build()
...
wfAgent, err := agentworkflow.NewAgent(wf, agentworkflow.AgentConfig{
    IncludeOutputsInResponse: true,
    Config:                   agent.Config{Name: "BilingualWorkflowAgent"},
})

NewAgent returns an ordinary *agent.Agent. Everything downstream — CreateSession, RunText, agent.Stream(true) — is the same API you use on a single leaf agent.

What to notice

• IncludeOutputsInResponse: true is the crux. Without it, only ResponseUpdates emitted by hosted agents surface; with it, the workflow’s terminal OutputEvent payload (the merged French + English messages) is surfaced in the wrapping agent’s response stream. Leave it off and the combined answer would be silently dropped.
• Concurrent = fan-out / fan-in. NewConcurrentWorkflowBuilder(french, english) dispatches the same input to both agents in parallel and aggregates their outputs into one batch. Swap in a sequential builder and the same wrapping code turns it into a pipeline instead — the caller never notices.
• Sessions keep the run alive. The first RunText opens a streaming workflow run; passing the same agent.WithSession(session) on the next turn reuses it rather than starting fresh:

session, _ := wfAgent.CreateSession(ctx)
for _, input := range prompts {
    for update, err := range wfAgent.RunText(ctx, input, agent.WithSession(session), agent.Stream(true)) {
        demo.Print(update, err)
    }
}

How it maps to the Agent Framework Go SDK

agentworkflow.NewAgent is the SDK seam that makes workflows composable as units: a graph becomes indistinguishable from a leaf agent, so multi-agent subsystems nest cleanly. This is the same wrapping used later in the observability lessons — here it’s the minimal, model-backed introduction to the pattern.

Run it

go test ./tutorial/03-workflows/agents/workflow_as_an_agent            # offline
AF_LIVE=1 go test ./tutorial/03-workflows/agents/workflow_as_an_agent  # + live (needs az login)
go run  ./tutorial/03-workflows/agents/workflow_as_an_agent            # needs Foundry

The offline test builds the workflow + wrapping agent with a fake credential and confirms the structure constructs and CreateSession succeeds with no network; the live model call is gated behind AF_LIVE=1.

Next: checkpoint_and_rehydrate