Do we have workflow (dynamic) as a subagent ?

[iincognitioo]

Hi

Requirements:

1. A user interacting with a chat agent -- Greetings, the agent explains what it does (e.g. create a word document), agent asks few questions (mandatory/optional). Then handover the task to say a workflow agent.
2. Workflow Agent
   a) Outline & HITL Nodes: Presents the document outline for approval from the user (HITL), Apply feedback iteratively (if any)..
   b) Code Agent node: Writes code & generates document and seek user feedback (HITL node)..Apply feedback (if any) .
   c) Returns to Main Agent ...

How can we implement this kind of pattern in ADK 2.0 especially with dynamic workflows.. ?
Anyone kindly point to any official sample or docs for this ?

Also, i see TASK API mentioned for google ADK 2.0:

Task API: Structured agent-to-agent delegation with multi-turn task mode, single-turn controlled output, mixed delegation patterns, human-in-the-loop, and task agents as workflow nodes.

Any docs on this ?

Thanks.

[surajksharma07]

@iincognitioo Yes, this pattern is fully supported in ADK 2.0 — you can use a Workflow as a sub-agent inside an Agent (chat agent) and the LLM hands off via transfer_to_agent.

For the HITL nodes, RequestInput inside a @node(rerun_on_resume=True) function handles the iterative approval loop cleanly, and dynamic dispatch to your code agent can be done via ctx.run_node().

One thing to note on the Task API: mode='task' agents currently can't be static graph nodes, so dispatch them via ctx.run_node() from a function node instead.

Good starting points in the adk-python repo: contributing/samples/workflows/dynamic_nodes/, request_input_advanced/, and patterns/workflow_triage/.

Could you try wiring these together and let us know if you hit any specific blockers?

[iincognitioo]

@node(rerun_on_resume=True)
async def test_workflow_node(ctx: Context, node_input: Any):
    yield Event(message="✅ Dynamic workflow subagent is working!")


test_workflow = Workflow(
    name="test_workflow",
    edges=[("START", test_workflow_node)],
)

root_agent = Agent(
    name="test_root_agent",
    model=os.environ.get("GEMINI_MODEL", "gemini-2.5-flash"),
    instruction="""
You are a test agent. When the user says 'test', 
transfer to test_workflow immediately.
""",
    sub_agents=[test_workflow],
)

Errors:
Input should be a valid dictionary or instance of BaseAgent [input_type=Workflow]

Workflow is NOT a BaseAgent subclass in ADK 2.0. Pydantic rejects it outright when you try to put it in sub_agents. The pattern is not supported at the framework level — sub_agents only accepts BaseAgent instances, and Workflow is a separate class hierarchy entirely ?

[surajksharma07]

@iincognitioo
The correct structure in ADK 2.0 is to invert the hierarchy: make Workflow the root (root_agent = Workflow(...)) and implement the conversation and requirement-gathering phase as a @node(rerun_on_resume=True) function that uses RequestInput to pause for user input, then dispatches the outline and code agents via ctx.run_node().
The ADK runner works with a Workflow root just as well as an Agent.

Two samples that map closely to your use case:

contributing/samples/workflows/request_input_rerun/agent.py covers the HITL draft → approve/revise → send loop with conditional routing,and
dynamic_nodes/agent.py shows the ctx.run_node() pattern for dispatching sub-agents dynamically.

Let us know if the restructured approach hits any blockers!