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

	Facebook
	Twitter
	Linkedin
	E-Mail

Generative AI > Google Gemini API > ADK Agent Routing - Intent-Based Dispatch

ADK Agent Routing - Intent-Based Dispatch

Author: Venkata Sudhakar

As you build more specialist agents, you need a way to route incoming requests to the right one. A routing agent classifies the user intent and delegates to the appropriate specialist. ShopMax India has 6 specialist agents - orders, returns, products, payments, technical support, and promotions. A single routing agent handles all incoming queries and dispatches them correctly.

The router uses a lightweight model for classification (fast and cheap) and then invokes the specialist agent. The classification step extracts intent, confidence, and any parameters the specialist needs from the original message.

The below example shows how ShopMax India implements intent-based routing across their specialist agent fleet.

from google.adk.agents import Agent
from google.adk.runners import Runner
from google.adk.sessions import InMemorySessionService
from google.adk.types import Content, Part
import json

# Define specialist agents
agents = {
    "orders": Agent(
        name="orders_agent", model="gemini-2.0-flash",
        instruction="You handle order tracking and status queries for ShopMax India."
    ),
    "returns": Agent(
        name="returns_agent", model="gemini-2.0-flash",
        instruction="You handle product returns and refund requests for ShopMax India."
    ),
    "products": Agent(
        name="products_agent", model="gemini-2.0-flash",
        instruction="You answer product questions and recommendations for ShopMax India."
    ),
    "payments": Agent(
        name="payments_agent", model="gemini-2.0-flash",
        instruction="You handle payment, EMI, and billing queries for ShopMax India."
    ),
}

# Lightweight router agent
router = Agent(
    name="router",
    model="gemini-2.0-flash",
    instruction="""Classify the customer query into one of: orders, returns, products, payments, general.
Respond with JSON only: {"intent": "<category>", "confidence": 0.95, "summary": "<brief rewrite>"}"""
)

The routing and dispatch logic,

session_svc = InMemorySessionService()

def route_and_respond(user_message: str, user_id: str) -> str:
    # Step 1: classify intent
    router_runner = Runner(agent=router, session_service=session_svc, app_name="shopmax")
    sess = session_svc.create_session(app_name="shopmax", user_id=user_id)
    result = router_runner.run(session_id=sess.id, user_id=user_id,
                               new_message=Content(parts=[Part(text=user_message)]))
    classification = ""
    for event in result:
        if event.is_final_response():
            classification = event.content.parts[0].text

routing = json.loads(classification)
    intent = routing["intent"]
    print(f"Routed to: {intent} (confidence: {routing['confidence']})")

# Step 2: dispatch to specialist
    target_agent = agents.get(intent, agents["products"])  # fallback to products
    specialist_runner = Runner(agent=target_agent, session_service=session_svc, app_name="shopmax")
    sess2 = session_svc.create_session(app_name="shopmax", user_id=user_id)
    result2 = specialist_runner.run(session_id=sess2.id, user_id=user_id,
                                    new_message=Content(parts=[Part(text=user_message)]))
    for event in result2:
        if event.is_final_response():
            return event.content.parts[0].text
    return ""

# Test routing
queries = [
    "Where is my order ORD-2024-88712?",
    "I want to return my headphones, they are defective",
    "What is the best laptop under Rs 50,000?"
]
for q in queries:
    print(f"Q: {q}")
    print(f"A: {route_and_respond(q, 'cust_123')[:80]}")
    print()

It gives the following output,

Q: Where is my order ORD-2024-88712?
Routed to: orders (confidence: 0.98)
A: I will look up order ORD-2024-88712 for you right away...

Q: I want to return my headphones, they are defective
Routed to: returns (confidence: 0.96)
A: I can help you with your defective headphones return...

Q: What is the best laptop under Rs 50,000?
Routed to: products (confidence: 0.99)
A: For Rs 50,000 budget I recommend the ASUS VivoBook...

For production at ShopMax India, log every routing decision to BigQuery with intent, confidence, user ID, and response time. This data shows which specialist agents are overloaded and where routing accuracy can be improved by refining the router instruction.

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