Asking the user¶

Set elicitation=True and the agent can stop to ask the user structured questions: a message plus a JSON Schema describing the answer, rendered by your UI as a form. This is ADK's native request_input tool. The package adds it to the root agent, makes the App resumable, and gives you helpers for the client side.

app = create_deep_agent(
    name="deployer",
    model_client="gemini-3-flash-preview",
    instruction="Confirm the target environment before deploying.",
    elicitation=True,
)

The flow¶

1. The agent asks. It calls adk_request_input(message, response_schema). ADK treats it as a long-running tool and stops the run. No further model call is made while it waits.
2. Your client shows the question. project_elicitation_event(event) returns an ElicitationRequest for the event that carries the question, and None for every other event.
3. The user answers. create_elicitation_response(request, answers) builds the user-role FunctionResponse with the same call ID.
4. The run resumes. Send that response with the interrupted invocation_id. ADK pairs it with the waiting call and the model receives {"result": answers}.

from adk_agent_harness import create_elicitation_response, project_elicitation_event

pending = None
async for event in runner.run_async(user_id=uid, session_id=sid, new_message=msg):
    if (request := project_elicitation_event(event)) is not None:
        pending = request

if pending is not None:
    # Render pending.message and pending.response_schema, collect answers.
    answers = {"environment": "staging", "notify_team": True}
    async for event in runner.run_async(
        user_id=uid,
        session_id=sid,
        invocation_id=pending.invocation_id,
        new_message=create_elicitation_response(pending, answers),
    ):
        ...

ElicitationRequest fields: id (call ID), message, response_schema, invocation_id, author (the agent that asked), and raw_event.

Schemas the model is told to use¶

The root instruction tells the model to ask once with an object schema covering all related questions, using standard JSON Schema: title and description for labels, enum for choices, boolean for confirmations, arrays for multi-select, and required for mandatory answers.

{
  "type": "object",
  "properties": {
    "environment": {"type": "string", "title": "Environment", "enum": ["staging", "production"]},
    "notify_team": {"type": "boolean", "title": "Notify the team?"}
  },
  "required": ["environment"]
}

Who can ask¶

If the user types instead of answering¶

If the next message is plain text rather than the FunctionResponse, the question is left without an answer. ToolCallRepairPlugin marks it as not completed before the next model call, so the model sees that the question went unanswered alongside the user's new message, and the provider doesn't reject the request. Your UI should still guide users to answer the form.