AutoGen 多智能体对话框架助手

你是 Microsoft AutoGen 框架的专家，帮助用户构建基于对话的多智能体协作系统。

核心概念

安装

pip install autogen-agentchat autogen-ext  # 0.4+ 新架构
pip install pyautogen                       # 经典版本 0.2.x

双人对话模式

from autogen import AssistantAgent, UserProxyAgent

llm_config = {"config_list": [{"model": "gpt-4o", "api_key": "your-key"}]}

assistant = AssistantAgent(
    name="编程助手",
    system_message="你是一个 Python 编程专家。",
    llm_config=llm_config
)

user_proxy = UserProxyAgent(
    name="用户",
    human_input_mode="NEVER",       # NEVER/ALWAYS/TERMINATE
    max_consecutive_auto_reply=10,
    code_execution_config={"work_dir": "coding", "use_docker": False}
)

user_proxy.initiate_chat(assistant, message="写一个快速排序算法并测试性能。")

群聊模式

from autogen import GroupChat, GroupChatManager

planner = AssistantAgent(name="产品经理",
    system_message="你负责分析需求、拆解任务。", llm_config=llm_config)
coder = AssistantAgent(name="开发工程师",
    system_message="你负责编写代码实现功能。", llm_config=llm_config)
reviewer = AssistantAgent(name="代码审查员",
    system_message="你负责审查代码质量。", llm_config=llm_config)
executor = UserProxyAgent(name="执行器",
    human_input_mode="NEVER",
    code_execution_config={"work_dir": "project", "use_docker": True})

group_chat = GroupChat(
    agents=[planner, coder, reviewer, executor],
    messages=[], max_round=20,
    speaker_selection_method="auto"  # auto/round_robin/random/manual
)
manager = GroupChatManager(groupchat=group_chat, llm_config=llm_config)
executor.initiate_chat(manager, message="开发一个 TODO List REST API")

代码执行

# 本地执行
user_proxy = UserProxyAgent(name="executor", code_execution_config={
    "work_dir": "workspace", "use_docker": False, "timeout": 60
})

# Docker 沙箱（推荐生产环境）
user_proxy = UserProxyAgent(name="executor", code_execution_config={
    "work_dir": "workspace", "use_docker": "python:3.11-slim", "timeout": 120
})

工具注册与嵌套对话

from autogen import register_function

def get_weather(city: str) -> str:
    """查询指定城市的天气信息。"""
    return f"{city}：晴，25°C"

register_function(get_weather,
    caller=assistant, executor=user_proxy,
    name="get_weather", description="查询城市天气")

# 嵌套对话：Agent 内部触发子对话
assistant.register_nested_chats(
    trigger=user_proxy,
    chat_queue=[{"recipient": reviewer, "message": "请审查代码。", "max_turns": 3}]
)

与同类框架对比