Skip to content

Day 2 - Module 6: Exercises

These exercises are based on the concepts and code presented in Module 6: Human-in-the-Loop Patterns (src/06-human-in-the-loop/).

Exercise 6.1: Changing the Interrupt Message

Goal: Practice modifying the information presented to the user when a workflow is interrupted.

  1. Open the src/06-human-in-the-loop/interrupt.py script.
  2. Locate the node2 function.
  3. Find the line where interrupt() is called: 
    human_input = await interrupt({"ai_answer": "What is your age?"})
  4. Change the dictionary passed to interrupt. For example, change the question or add more context: 
    human_input = await interrupt({"prompt_for_user": "Please provide your favorite color to continue.", "current_step": "node2"})
  5. Run the script (python interrupt.py).
  6. Observe the output when the script pauses. Does it display the new message or data structure you provided in the interrupt call?

Exercise 6.2: Modifying the Reviewer Agent

Goal: Observe how changing an agent's instructions affects its behavior in a collaborative loop.

  1. Open the src/06-human-in-the-loop/report-agents.py script.
  2. Part A: Change Initial Input:
    • Locate the main function and the initial user prompt: 
      user_input = input("User: ")
    • Run the script (python report-agents.py). When prompted for User:, provide a different piece of text than the example, e.g., "Write a tweet about the importance of testing software."
    • Observe the interaction between the Writer and Reviewer agents.
  3. Part B: Modify Reviewer Instructions:
    • Stop the script if it's running.
    • Find the REVIEWER_INSTRUCTIONS variable.
    • Modify the instructions to make the reviewer stricter or more lenient. For example:
      • Stricter: Add a requirement like "Ensure the text is less than 100 characters and includes a hashtag."
      • Lenient: Change the instructions to something like "Briefly check for major grammatical errors only. Approve if it looks reasonable."
    • Run the script again, providing the same initial input as in Part A.
    • Compare the interaction to Part A. Does the reviewer's feedback change? Does the conversation take more or fewer turns to complete?

Exercise 6.3 (Conceptual): Modifying Interrupt for Approval

Goal: Think about how to adapt the interrupt pattern for a yes/no approval step.

  1. Consider the interrupt.py example where node2 interrupts to get the user's age.
  2. Imagine you want node2 to present some data (e.g., result = "Plan A") and ask the human for approval before node3 runs.
  3. How would you modify node2?
    • What dictionary would you pass to interrupt()? (e.g., {"data_to_approve": result, "prompt": "Do you approve this plan? (yes/no)"})
    • How would you check the human_input received after the interrupt?
  4. How would you modify the graph structure (builder.add_edge)? Conditional edges would be needed after node2.
    • If human_input is "yes", add an edge from node2 to node3.
    • If human_input is "no", what should happen? Add an edge back to node1? Add an edge to a new "handle_rejection" node? Add an edge directly to END?