Skip to main content

Lab 21.7 Solution: Peer-to-Peer Agent Hand-offs

Goal​

This file contains the complete code for the agent.py script using ADK 2.0 Collaborative Workflows.

collaborative_duo/agent.py​

from google.adk import Agent

# 1. Define the Tech Agent
# We define it without collaborators initially to avoid reference errors.
tech_agent = Agent(
    name="tech_agent",
    model="gemini-3.5-flash",
    description="Expert in hardware specs and troubleshooting.",
    instruction="""
    You are a technical support expert. Your job is to answer deep technical questions.
    
    **HAND-OFF RULES:**
    - If the user asks about prices, discounts, or expressed intent to BUY, 
      you MUST hand off to the sales_agent.
    """
)

# 2. Define the Sales Agent
# This agent already knows about the tech_agent.
sales_agent = Agent(
    name="sales_agent",
    model="gemini-3.5-flash",
    description="Expert in pricing, bundles, and purchase orders.",
    instruction="""
    You are a sales representative. Your job is to help users with pricing and buying.
    
    **HAND-OFF RULES:**
    - If the user asks about hardware specs, compatibility, or technical issues, 
      you MUST hand off to the tech_agent.
    """,
    collaborators=[tech_agent] 
)

# 3. Retroactively link the tech_agent back to sales_agent
# This enables bi-directional hand-offs.
tech_agent.collaborators = [sales_agent]

# 4. Set the Root Agent
# Any agent in the collaborative group can be the entry point.
root_agent = sales_agent

Self-Reflection Answers​

  1. What is the main difference between this and the Coordinator pattern?

    • Answer: In the Coordinator pattern, there is a "boss" agent that manages sub-agents. Control always returns to the coordinator after a sub-agent finishes. In a Collaborative Workflow, control stays with the new agent after a hand-off. It is peer-to-peer rather than hierarchical.

  2. How does the LLM know how to perform the hand-off?

    • Answer: The ADK automatically creates a tool (e.g., hand_off_to_tech_agent) for every agent listed in the collaborators parameter. The LLM sees these tools in its toolkit and calls them when its instructions tell it to delegate.

  3. Why did we have to set tech_agent.collaborators retroactively?

    • Answer: In Python, you cannot refer to a variable (sales_agent) before it has been defined. By setting the collaborators list on the second agent and then updating the first one, we create the circular reference required for bi-directional hand-offs.