Skip to content

Day 3 - Module 8: Exercises

These exercises are based on the concepts and code presented in Module 8: Society of Agents (AutoGen/MagenticOne) (src/08-society-of-agents/).

Exercise 8.1: Changing the Task (Simple Group)

Goal: Observe how the MagenticOne orchestrator directs different agents based on the initial task.

  1. Open the src/08-society-of-agents/simple-group.py script.
  2. Locate the line where the chat is initiated: 
    stream = magenticteam.run_stream(task="what time is it here?.")
  3. Change the initial task string to ask for different information that requires different agents. Examples:
    • task="Where does Dennis live?" (Should primarily involve users_agent and location_agent)
    • task="Who is the current user?" (Should primarily involve users_agent)
  4. Run the script (python simple-group.py).
  5. Observe the conversation flow printed to the console. Does the orchestrator correctly involve the agents relevant to the new task? Does the summary_agent still get involved appropriately?

Exercise 8.2: Adding a Weather Agent (Simple Group)

Goal: Practice adding a new agent with a specific capability to the group chat.

  1. Open the src/08-society-of-agents/simple-group.py script.
  2. Define a weather tool: Add a get_weather(location: str) function (you can make it return a fixed string like "It is sunny in [location]" for simplicity).
  3. Define a weather agent: Create a new AssistantAgent named weather_agent.
    • Give it a relevant description (e.g., "Knows the weather.").
    • Provide it with the get_weather tool function.
    • Use the same model_client as the other agents.
  4. Add agent to group: Include the weather_agent in the agents list passed to MagenticOneGroupChat.
  5. Modify the task: Change the initial task to include asking about the weather, e.g., task="What is the weather like where Dennis lives?".
  6. Run the script (python simple-group.py).
  7. Observe the conversation. Does the orchestrator involve the new weather_agent after finding Dennis's location? Does the final summary include the weather information?

Exercise 8.3: Testing Chef Agent Logic (Chef Group)

Goal: Verify if the chef agent correctly handles missing information (allergies) based on its instructions.

  1. Open the src/08-society-of-agents/chef-and-group.py script.
  2. Locate the get_medical_history tool function.
  3. Modify the function to return an empty string or a message indicating no known allergies: 
    def get_medical_history(username: str) -> str:
    """Returns the medical history for a given username."""
    if username == "Dennis":
        # return "Dennis is allergic to peanuts."
        return "No known allergies for Dennis."
    else:
        return "Unknown user."
  4. Run the script (python chef-and-group.py) with the task "I want to have something to eat. What would you recommend?."
  5. Carefully read the conversation flow. Does the chef_agent, after finding out there are no known allergies from the users_agent, explicitly ask something like "Do you have any allergies I should be aware of?" before recommending a dish? (This depends on its system prompt instructions being followed correctly by the LLM).

Exercise 8.4: Adding Constraints and preferences into a planning agent

Goal: Understand how a group of agents can collaborate without a moderator. Learn how context agents (location, user preferences) feed into a planning agent (chef). Work with a real agent configuration and extend it with constraints.

  1. Add Domain Expertise (10–15 min):

  2. Add a new agent: nutritionist_agent.

    • Role: Evaluate the nutritional quality of the suggested meal.
    • Provide it a tool: analyze_nutrition(meal: str) — mock this as a function returning “balanced”, “high-fat”, etc.

  3. Inject a Constraint (5 min):

  4. Modify users_agent to include a dietary preference, e.g., "low-carb" or "vegan".

  5. Ensure the chef_agent respects this constraint when generating meal options.

  6. Optional Extensions:

  7. Add a budget_agent with a get_cost_estimate(meal) tool.

  8. Let agents negotiate (by reasoning in messages) over trade-offs between cost and nutrition.

Exercise 8.5: Modifying Reasoning Oversight (o1 Group)

Goal: Change the focus of the quality check performed by the reasoning agent.

  1. Open the src/08-society-of-agents/o1-with-chef-group.py script.
  2. Locate the check_conversation async function (which acts as a tool for the consultation_agent).
  3. Find the prompt string passed to reasoning_agent.on_messages inside this function: 
    prompt = "Check the messages for inconsistencies or open questions..."
  4. Modify this prompt to ask the reasoning_agent (o1-mini) to check for something specific. Examples:
    • prompt = "Review the conversation. Did the chef confirm the user's allergies before recommending a meal? Answer yes or no and explain briefly."
    • prompt = "Analyze the conversation. Was the final meal recommendation appropriate given the available ingredients mentioned earlier?"
  5. Run the script (python o1-with-chef-group.py).
  6. Observe the output from the consultation_agent (which includes the response from the reasoning_agent). Does the feedback now focus specifically on the new check you asked the reasoning_agent to perform?