Day 3 - Module 8: Solutions

These are the solutions or discussion points for the exercises in Module 8.

Solution 8.1: Changing the Task (Simple Group)

Modify src/08-society-of-agents/simple-group.py:

# ... (imports, model client, tools, agents definition remain the same) ...

if __name__ == "__main__":
    # ... (group chat setup) ...

    # --- Modified Task --- 
    task = "Where does Dennis live?"
    # -------------------

    stream = magenticteam.run_stream(task=task)
    console = Console()
    console.print(stream)

Expected Output: The conversation flow should primarily involve the users_agent (to confirm the user is Dennis) and the location_agent (to get the location for Dennis). The time_agent should not be involved. The summary_agent is called at the end to state the final answer (e.g., "Dennis lives in Berlin.") before terminating.

Solution 8.2: Adding a Weather Agent (Simple Group)

Modify src/08-society-of-agents/simple-group.py:

# ... (imports, model client, other tools definitions remain the same) ...

# --- Define New Weather Tool --- 
def get_weather(location: str) -> str:
    """Returns the current weather for a given location."""
    print(f"Tool: Getting weather for {location}")
    # Simple mock response
    return f"It is currently sunny in {location}."
# -----------------------------

# ... (definitions for users_agent, location_agent, time_agent, summary_agent) ...

# --- Define New Weather Agent --- 
weather_agent = AssistantAgent(
    name="WeatherAgent",
    description="Knows the weather.",
    system_message="You provide weather information.",
    model_client=model_client,
    tools=[get_weather]
)
# ------------------------------

if __name__ == "__main__":
    # --- Add weather_agent to the list --- 
    agents = [users_agent, location_agent, time_agent, summary_agent, weather_agent]
    # -------------------------------------

    magenticteam = MagenticOneGroupChat(
        agents=agents,
        model_client=model_client, # Orchestrator model
        termination_condition=MaxMessageTermination(10) # Increased limit for more steps
    )

    # --- Modified Task --- 
    task = "What is the weather like where Dennis lives?"
    # -------------------

    stream = magenticteam.run_stream(task=task)
    console = Console()
    console.print(stream)

Expected Output: The conversation should involve: 1. users_agent (confirming user is Dennis). 2. location_agent (getting Dennis's location, e.g., Berlin). 3. weather_agent (being called with the location "Berlin" to get the weather). 4. summary_agent (providing the final answer, e.g., "The weather where Dennis lives (Berlin) is currently sunny.") before terminating.

Solution 8.3: Testing Chef Agent Logic (Chef Group)

Modify src/08-society-of-agents/chef-and-group.py:

# ... (imports, model client, other tools definitions remain the same) ...

# --- Modified Medical History Tool --- 
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."
# -----------------------------------

# ... (agent definitions, group chat setup, task remain the same) ...

Expected Output: When the conversation reaches the chef_agent, having already learned the username is Dennis but that there are "No known allergies", the chef_agent asks a clarifying question based on its system prompt (which includes instructions like "If allergies are unknown, ask the user"). The output should show the chef_agent asking something like:

ChefAgent: Okay Dennis, I see no known allergies on file. To make sure I recommend something safe, do you have any food allergies or dietary restrictions I should be aware of?

This demonstrates the agent following its instructions to proactively seek missing critical information.

Exercise 8.4: Adding Constraints and preferences into a planning agent

File: chef-and-group.py Concepts: Contextual agents, collaboration, basic tool use

🔧 Expected Code/Changes

Added `nutritionist_agent`

def analyze_nutrition(meal: str) -> str:
    if "pasta" in meal.lower():
        return "High in carbs."
    if "burger" in meal.lower():
        return "High in fat"
    return "Well-balanced."

nutritionist_agent = AssistantAgent(
    "Nutritionist",
    model_client=model_client,
    tools=[analyze_nutrition],
    description="A helpful assistent that Evaluates the nutritional of meals.",
)

Updated user_agent with dietary constraint

async def get_diet_preference(username: str) -> str:
    "Get the diet preference for a given username."
    print("executing get_diet_preference")
    return f"{username} is a vegetarian and does not eat meat."

Updated group to include new agent

group = MagenticGroupChat(
    agents=[
        user_agent,
        location_agent,
        time_agent,
        chef_agent,
        nutritionist_agent,
        summary_agent
    ]
)

Solution Behavior/Expected Output:

The chef_agent should now consider "low-carb" preferences and avoid recommending dishes like pasta.
The nutritionist_agent should evaluate the dish, e.g.:

“This dish is balanced and aligns with the user’s low-carb goal.”

Solution 8.4: Modifying Reasoning Oversight (o1 Group)

Modify src/08-society-of-agents/o1-with-chef-group.py:

# ... (imports, model clients, tools, agent definitions remain the same) ...

# Tool for consultation agent to check conversation quality
async def check_conversation(messages: list[dict[str, str]]) -> str:
    """Checks the conversation for inconsistencies or open questions."""
    # --- Modified Prompt for Reasoning Agent --- 
    prompt = "Review the conversation. Did the chef confirm the user's allergies before recommending a meal? Answer yes or no and explain briefly."
    # -----------------------------------------
    response = await reasoning_agent.on_messages(messages=[ChatMessage(role="user", content=prompt)])
    return response[0].content

# ... (consultation_agent definition using check_conversation tool) ...
# ... (group chat setup, task, main execution remain the same) ...

Expected Output: When the consultation_agent is invoked by the orchestrator, its output (which comes from the reasoning_agent via the check_conversation tool) should now specifically address whether the chef confirmed allergies. For example:

If the chef did ask: ConsultationAgent: Yes, the chef asked about allergies after learning none were on file before proceeding with the recommendation.
If the chef did not ask (due to an LLM error or prompt issue): ConsultationAgent: No, the chef did not explicitly confirm allergies with the user before making a recommendation, even though the initial medical history check showed no known allergies.

This demonstrates tailoring the oversight provided by the reasoning agent to focus on specific quality aspects of the multi-agent interaction.