Multi Agent Framework

Introduction

Multi-Agent Framework build from scratch, based on OpenAI API, to enable multiple agents to talk to each other, perform tasks, and handle basic control with a human in the loop.

You can also add third-party agents like Dify to the group, and the framework will automatically handle the handoff between agents.

It is built to be flexible and user-friendly, enabling users to create and manage groups of agents with various roles and skills. This framework can be applied in many areas, such as customer support, research, and education.

Simple Demo

Try it: Simple demo for multi-agent

Tutorial

• Simpel demo for single agent
• How to build Group of Agents
• Chat with Group of Agents
• Internal dialogue within Group of Agents
• Task for Group of Agents
• Low Level API for Group Discussion with Human in the Loop

Usage

Pre-requirements

pip install -r requirements.txt
cp .env_example .env

write your own KEY and URL(optional)

OPENAI_API_KEY=xxx
OPENAI_BASE_URL=xxx

Step Zero

from dotenv import load_dotenv
from openai import OpenAI
from src import Env,Agent,Group

load_dotenv()
model_client = OpenAI()

Step One

Agent is the basic unit of the framework, it can build from scratch or connect to third-party agents

Creat Agent like this

artist = Agent(name="artist",
        role="Artist", 
        description="Transfer to me if you need help with art.",
        persona = "You are a professional artist who has been working in the industry for over 10 years. You have a deep understanding of art history and have a strong passion for creating art. You are known for your unique style and innovative approach to art. You are always looking for new ways to express yourself and push the boundaries of what is possible in the art world.",
        model_client=model_client,
        verbose=True)

can use agent like this

response = artist.do("Can you help me with art?",model="gpt-4o-mini")

can add tools like this

def web_search(qury:str)->str:
    """
    web search tool
    """
    # do web search
    return "web search result"

researcher = Agent(name="researcher",
        role="Researcher",
        description="Transfer to me if you need help with research.",
        persona = "You are a professional researcher who can do web search to conduct research on a wide range of topics. You have a deep understanding of how to find and evaluate information from a variety of sources. You are known for your ability to quickly find relevant information and present it in a clear and concise manner.",
        tools=[web_search],
        model_client=model_client,
        verbose=True)

or equip with memory like this

telos = Agent(name="telos",
              role="Assistant",
              description="Transfer to me if you need help with general questions.",
              persona="You are a general assistant who can help with a wide range of questions. You have a deep understanding of a variety of topics and can provide information and assistance on a wide range of subjects. You are known for your ability to quickly find answers to questions and provide helpful information in a clear and concise manner.",
              model_client=model_client,
              verbose=True)

telos.init_memory(working_memory_threshold=3)

or connect a third-party agent that was created at Dify like this.

mathematician = Agent(name="mathematician",
    role="Mathematician", 
    description="Transfer to me if you need help with math.", 
    dify_access_token="app-rlK8IzzWCVkNbkxxxxxxx",
    verbose=True)
# persona is not needed for Dify agent, it already has its own persona

or connect a websocket agent like this.

agent = Agent(name="assistant", 
              role="Assistant",
              description="Transfer to me if you need help",
              websocket_url="ws://localhost:5358/ws_agent_demo",
              verbose=True)

Step Two

Env is the environment where agents live, you can add a description and agents to the environment. In addition,it can be created with or without relationships between agents, and can also set the language used in the environment. Env will be used to create a group of agents.

Create Env like this (all agents are fully connected by default)

env = Env(
    description="This is a test environment",
    members=[mathematician, artist]
)

or like this (self-defined topology relationships between agents)

env = Env(
    description="This is a test environment",
    members=[mathematician, artist],
    relationships={"agent1": ["agent2"]}
)

or set language used in the environment

env = Env(
    description="This is a test environment",
    members=[mathematician, artist],
    language="中文"
)

Step Three

Group is a collection of agents that can be used to chat, perform tasks, and handle basic control with a human in the loop.

Build Group like this

g = Group(env=env,model_client=model_client,verbose=True)

can add extra agent into group dynamically like this

designer = Agent(name="designer",
    role="Designer", 
    description="Transfer to me if you need help with design.", 
    model_client=OpenAI(),
    verbose=True)

g.add_member(designer)

or delete agent from group dynamically like this

takeaway,observed_speakers = g.delete_member("artist") # delete by name
# will return takeaway and observed_speakers for memory retrieval in the future

or invite agent to join group dynamically like this

# automatically create agent
g.invite_member("a philosopher who calls himself agent4 , he is a big fan of plato and aristotle")

or dismiss the group like this

g.dismiss_group()
# when the group is dismissed, all agents will be deleted and each of them will get their own memory back

Step Four

Some examples of how to use the group

chat with group of agents(dynamic agent selection)

response= g.chat("Can you explain the concept of complex numbers?",model="gpt-4o-mini")
response= g.chat("Can you help me with art?",model="gpt-4o-mini")

internal dialogue within group of agents based on the current environment description

g.dialogue(model="gpt-4o-mini",max_turns=10)

task for group of agents

response = g.task("I want to build a simplistic and user-friendly bicycle help write a design brief.",model="gpt-4o-mini",strategy="auto")

low-level API example

g.user_input("can you help me with math?")
next_agent = g.handoff(next_speaker_select_mode="auto",include_current=True,model="gpt-4o-mini")

g.user_input("Discuss the concept of abstract art.")
response = g.call_agent(next_speaker_select_mode="auto",include_current=True,model="gpt-4o-mini")
response = g.call_agent(next_speaker_select_mode="auto",include_current=True,model="gpt-4o-mini")
g.user_input("how do you feel about abstract art?")
response = g.call_agent(next_speaker_select_mode="auto",include_current=True,model="gpt-4o-mini")
response = g.call_agent(next_speaker_select_mode="auto",include_current=True,model="gpt-4o-mini")
response = g.call_agent(next_speaker_select_mode="auto",include_current=True,model="gpt-4o-mini")