This sample demonstrates a workflow that incorporates human-in-the-loop, where the agent can pause and ask for user input before continuing.

The architecture of this sample is a WorkflowAgent that acts as an interactive concierge to create an itinerary.

• initial_prompt: A Python function that starts the workflow by showing an introductory message and asking for user input.
• concierge_agent: An LlmAgent that takes the user's input and generates a list of activities.
• get_user_feedback: A Python function that shows the generated itinerary and asks the user for feedback.
• process_feedback: A Python function that takes the user's feedback and sends it back to the concierge_agent.

The workflow is defined with a loop, allowing the user to iteratively refine the itinerary:

edges=[
    (
        "START", 
        initial_prompt, 
        concierge_agent, 
        get_user_feedback, 
        process_feedback
    ),
    (process_feedback, concierge_agent)
]

This sample showcases the Human-in-the-Loop (HITL) capability of WorkflowAgent. The RequestInput event is used to pause the workflow and request input from the user.

• RequestInput: This event is yielded by a function node to send a message to the user and wait for a response.

This feature is essential for building interactive applications where user feedback or decisions are required to drive the workflow. The sample demonstrates a conversational loop where the agent and user can go back and forth to achieve a desired outcome.

To deploy this workflow agent, you can use the adk deploy command.

Ensure you have authenticated with Google Cloud:

gcloud auth application-default login

Your GCP project and location should be set in a .env file in the root of this project.

adk deploy workflows-HITL/agent.py:root_agent --display-name "Interactive Concierge Agent"

After deployment, you can interact with the agent through the provided endpoint. The agent will prompt you for input and guide you through creating an itinerary.