Skip to main content

Lab 19 Solution: Building a Collaborative Travel Team

Goal​

This file contains the complete code for the agent.py script using ADK 2.0 Collaboration Modes.

travel_team/agent.py​

from google.adk import Agent

# 1. Define the Weather Specialist
# We use 'single_turn' because this is a quick data retrieval task
# with no need for further user interaction.
weather_agent = Agent(
    name="weather_checker",
    model="gemini-3.5-flash",
    mode="single_turn",
    # Required on every agent in a sub_agents dispatch chain -- without it:
    # "ValueError: A node must have rerun_on_resume=True."
    rerun_on_resume=True,
    instruction="Provide a brief, enthusiastic 3-day weather forecast for the user's destination."
)

# 2. Define the Flight Booker
# We use 'task' mode because the agent might need to ask 
# clarification questions before it considers its task "complete".
flight_agent = Agent(
    name="flight_booker",
    model="gemini-3.5-flash",
    mode="task",
    # Required for any node that might pause and resume across turns --
    # without it: "ValueError: A node must have rerun_on_resume=True."
    rerun_on_resume=True,
    instruction="""
    Help the user book a flight. 
    1. Ask for their preferred airline or time if not provided.
    2. Once you have the info, confirm the 'booking' (simulated) and stop.
    """
)

# 3. Define the Coordinator
# The root agent manages the high-level flow. 
# It doesn't need a mode.
root_agent = Agent(
    name="travel_planner",
    model="gemini-3.5-flash",
    # The coordinator itself also needs this -- it's part of the same
    # dispatch chain and can be "woken up" when a sub-agent's task resumes.
    rerun_on_resume=True,
    instruction="""
    You are a travel planning coordinator.
    Your goal is to build a complete plan for the user.
    
    PROCESS:
    1. Call the `weather_checker` to get the forecast.
    2. Call the `flight_booker` to arrange travel.
    3. Once both sub-tasks are done, present a final summary to the user.
    """,
    sub_agents=[weather_agent, flight_agent] 
)

Self-Reflection Answers​

  1. Why would you use single_turn instead of task for a database lookup node?

    • Answer: Efficiency. single_turn is faster and cheaper because it forbids any further interaction. If a node's job is purely to fetch data and return it, single_turn ensures it doesn't accidentally start a conversation with the user, which would waste tokens and time.

  2. What happens to the conversation history when a sub-agent is in task mode?

    • Answer: In ADK 2.0, task-mode agents operate in a "branch" of the session. While they have access to the relevant history to perform their task, the parent agent remains the "source of truth." When the task completes, the results are bubbled up to the parent's context.

  3. How does the mode setting improve the reliability of complex workflows?

    • Answer: It makes the "Return to Parent" behavior deterministic. Instead of hoping the LLM will remember to hand back control, the framework enforces the return based on the mode. This prevents "lost in thought" scenarios where a sub-agent continues chatting with the user indefinitely instead of returning to the main workflow.