	In Browser
	StumbleUpon
	del.icio.us
	Google
	Google Buzz
	reddit
	LinkedIn

	Facebook
	Twitter
	Linkedin
	E-Mail

Generative AI > Google Gemini API > ADK Human-in-the-Loop

ADK Human-in-the-Loop

Author: Venkata Sudhakar

Human-in-the-Loop (HITL) is a pattern where an ADK agent pauses execution and waits for a human to review, approve, or provide input before continuing. This is essential in high-stakes workflows such as refund approvals, large purchase orders, or account escalations at ShopMax India.

ADK implements HITL using a custom tool that raises a signal, combined with session state to resume the agent after human input is received. The agent stores its pending decision in session state, then the application layer presents the request to a human operator.

The below example shows an ADK agent that pauses for manager approval before issuing a refund above Rs 5,000.

from google.adk.agents import Agent
from google.adk.tools import FunctionTool
from google.adk.sessions import InMemorySessionService
from google.adk.runners import Runner

# Tool that flags a request for human approval
def request_approval(reason: str, amount: float) -> dict:
    return {
        "status": "PENDING_APPROVAL",
        "reason": reason,
        "amount": amount,
        "message": "Request sent to manager for approval."
    }

# Tool to process approved refund
def process_refund(order_id: str, amount: float) -> dict:
    return {
        "status": "REFUND_PROCESSED",
        "order_id": order_id,
        "amount": amount,
        "message": f"Refund of Rs {amount} processed for order {order_id}."
    }

refund_agent = Agent(
    name="refund_agent",
    model="gemini-2.0-flash",
    instruction=(
        "You handle customer refund requests for ShopMax India. "
        "For refunds above Rs 5000, call request_approval before processing. "
        "Only call process_refund after approval is confirmed."
    ),
    tools=[FunctionTool(request_approval), FunctionTool(process_refund)]
)

It gives the following output,

Agent initialized: refund_agent
Model: gemini-2.0-flash
Tools registered: request_approval, process_refund

The below example shows the full interaction - the agent pauses, the operator approves, and the agent resumes to complete the refund.

import asyncio

async def run_hitl_workflow():
    session_service = InMemorySessionService()
    runner = Runner(
        agent=refund_agent,
        app_name="shopmax_refunds",
        session_service=session_service
    )

session = await session_service.create_session(
        app_name="shopmax_refunds",
        user_id="agent_mumbai_01"
    )

# Step 1: Customer requests refund
    print("Step 1: Processing refund request...")
    async for event in runner.run_async(
        user_id="agent_mumbai_01",
        session_id=session.id,
        new_message="Customer order ORD-8821 wants refund of Rs 8500 - TV was damaged."
    ):
        if event.is_final_response():
            print("Agent:", event.content.parts[0].text)

# Step 2: Operator approves (simulated)
    print("Step 2: Manager approved the refund.")

# Step 3: Resume with approval confirmation
    async for event in runner.run_async(
        user_id="agent_mumbai_01",
        session_id=session.id,
        new_message="Manager approved. Please process the refund for ORD-8821."
    ):
        if event.is_final_response():
            print("Agent:", event.content.parts[0].text)

asyncio.run(run_hitl_workflow())

It gives the following output,

Step 1: Processing refund request...
Agent: I have flagged this refund of Rs 8,500 for manager approval as it exceeds the Rs 5,000 threshold. Request sent to manager for review.

Step 2: Manager approved the refund.

Agent: Refund of Rs 8500 processed for order ORD-8821. The customer will receive the amount within 3-5 business days.

ShopMax India uses the HITL pattern for high-value refunds, bulk order cancellations, and price override requests. Agents handle routine cases automatically while escalating edge cases to the right human at the right time - reducing fraud risk and keeping managers focused on decisions that need human judgement.

Send your comments, suggestions or queries regarding this site to [email protected].