Skip to content

Three-Agent Collaboration

Script in langroid-examples

A full working example for the material in this section is in the three-agent-chat-num.py script in the langroid-examples repo: examples/quick-start/three-agent-chat-num.py.

Let us set up a simple numbers exercise between 3 agents. The Processor agent receives a list of numbers, and its goal is to apply a transformation to each number \(n\). However it does not know how to apply these transformations, and takes the help of two other agents to do so. Given a number \(n\),

  • The EvenHandler returns \(n/2\) if n is even, otherwise says DO-NOT-KNOW.
  • The OddHandler returns \(3n+1\) if n is odd, otherwise says DO-NOT-KNOW.

As before we first create a common ChatAgentConfig to use for all agents:

config = lr.ChatAgentConfig(
    llm = lr.language_models.OpenAIGPTConfig(
        chat_model=lr.language_models.OpenAIChatModel.GPT4,
    ),
    vecdb=None,
)

Next, set up the processor_agent, along with instructions for the task: 

processor_agent = lr.ChatAgent(config)
processor_task = lr.Task(
    processor_agent,
    name = "Processor",
    system_message="""
        You will receive a list of numbers from the user.
        Your goal is to apply a transformation to each number.
        However you do not know how to do this transformation,
        so the user will help you. 
        You can simply send the user each number FROM THE GIVEN LIST
        and the user will return the result 
        with the appropriate transformation applied.
        IMPORTANT: only send one number at a time, concisely, say nothing else.
        Once you have accomplished your goal, say DONE and show the result.
        Start by asking the user for the list of numbers.
        """,
    llm_delegate=True, #(1)!
    single_round=False, #(2)!
)

  1. Setting the llm_delegate option to True means that the processor_task is delegated to the LLM (as opposed to the User), in the sense that the LLM is the one "seeking" a response to the latest number. Specifically, this means that in the processor_task.step() when a sub-task returns DO-NOT-KNOW, it is not considered a valid response, and the search for a valid response continues to the next sub-task if any.
  2. single_round=False means that the processor_task should not terminate after a valid response from a responder.

Set up the other two agents and tasks:

NO_ANSWER = lr.utils.constants.NO_ANSWER

even_agent = lr.ChatAgent(config)
even_task = lr.Task(
    even_agent,
    name = "EvenHandler",
    system_message=f"""
    You will be given a number. 
    If it is even, divide by 2 and say the result, nothing else.
    If it is odd, say {NO_ANSWER}
    """,
    single_round=True,  # task done after 1 step() with valid response
)

odd_agent = lr.ChatAgent(config)
odd_task = lr.Task(
    odd_agent,
    name = "OddHandler",
    system_message=f"""
    You will be given a number n. 
    If it is odd, return (n*3+1), say nothing else. 
    If it is even, say {NO_ANSWER}
    """,
    single_round=True,  # task done after 1 step() with valid response
)

Now add the even_task and odd_task as subtasks of the processor_task, and then run it as before:

processor_task.add_sub_task([even_task, odd_task])
processor_task.run()

Feel free to try the working example script three-agent-chat-num.py langroid-examples repo: examples/quick-start/three-agent-chat-num.py:

python3 examples/quick-start/three-agent-chat-num.py

Here's a screenshot of what it looks like: three-agent-num.png

Next steps

In the next section you will learn how to use Langroid to equip a ChatAgent with tools or function-calling.