Routing Agents for Specific Functions - Multi-agent supervisor Concept

[mraguth]

I am referring to the concept described in this link, and I find it to be a brilliant idea. In my use case, I have three agents, each responsible for a different type of function. One agent handles 'math', another focuses on 'science', and the third specializes in 'research'.

My question is: How can I define these agents or provide them with specific information so that a supervisor can determine which agent to send a question to, instead of simply passing the question along? Any thoughts or suggestions would be greatly appreciated.

Link to the concept

[dosubot[bot]]

You can define and route these agents by creating instances of an Agent class for each specific function and using a supervisor to decide which agent to use based on the input. Here's a simplified example of how you might implement this:

# Create agent instances
math_agent = Agent.from_llm_and_tools(llm, math_tools, callback_manager, output_parser)
science_agent = Agent.from_llm_and_tools(llm, science_tools, callback_manager, output_parser)
research_agent = Agent.from_llm_and_tools(llm, research_tools, callback_manager, output_parser)

# Supervisor logic to decide which agent to use
def route_question(user_input):
    if user_input == 'math':
        action = math_agent.plan(intermediate_steps, callbacks, **kwargs)
    elif user_input == 'science':
        action = science_agent.plan(intermediate_steps, callbacks, **kwargs)
    elif user_input == 'research':
        action = research_agent.plan(intermediate_steps, callbacks, **kwargs)
    return action

In this example, llm, math_tools, science_tools, research_tools, callback_manager, output_parser, intermediate_steps, callbacks, and kwargs are placeholders that should be replaced with your actual objects and data. The plan method is used to determine the action based on the user input ^[1].

To continue talking to Dosu, mention @dosu.

[mraguth]

@dosu In this examples all agents are react agents. Can you change your concept to react agets? Is that possible?

[dosubot[bot]]

Yes, it's possible to modify the concept to use ReAct agents. You can create ReAct agents using the create_react_agent method from LangGraph, which allows for more flexibility and control. Here's how you can adapt the example to use ReAct agents:

from langgraph.prebuilt import create_react_agent
from langchain_openai import ChatOpenAI

# Define your language model
model = ChatOpenAI(model="gpt-4o")

# Define tools for each agent
math_tools = [/* your math tools */]
science_tools = [/* your science tools */]
research_tools = [/* your research tools */]

# Create ReAct agents
math_agent = create_react_agent(model, math_tools)
science_agent = create_react_agent(model, science_tools)
research_agent = create_react_agent(model, research_tools)

# Supervisor logic to decide which agent to use
def route_question(user_input):
    if user_input == 'math':
        messages = math_agent.invoke({"messages": [("human", user_input)]})
    elif user_input == 'science':
        messages = science_agent.invoke({"messages": [("human", user_input)]})
    elif user_input == 'research':
        messages = research_agent.invoke({"messages": [("human", user_input)]})
    return messages["messages"][-1].content

This setup uses ReAct agents, which are designed to synergize reasoning and acting in language models, providing a structured approach to handling tasks ^[1][2].

To continue talking to Dosu, mention @dosu.

[mraguth]

It is woking fine. However, if I switch from math to science right away, it is confusing and giving wrong answer or not able to see science tools. Is there way I initialize between these switches? Thoughts?