OpenAI多智能体框架Swarm实测—基于Qwen开源模型

1、基础配置

底层模型：Qwen开源模型
swarm：0.1.0

2、agent设置

agent_a = Agent(name="Agent A", instructions="你是Agent A，是一个有用的助手", functions=[transfer_to_agent_b], model=model_name,
                tool_choice='auto')
agent_b = Agent(name="Agent B", instructions="你是Agent B，仅仅使用繁体字说话", functions=[transfer_to_agent_a],model=model_name, tool_choice='auto')

agent之间的跳转函数

def transfer_to_agent_b():
    return agent_b
def transfer_to_agent_a():
    return agent_a

3、单智能体示例

输入示例：

msg = [{"role": "user", "content": "你好，小明的年龄是12岁。"},
       {"role": "user", "content": "我的年龄是小明的2倍"},
       {"role": "user", "content": "请问我的年龄是多少"}]

4、多智能体示例

输入示例

msg = [{"role": "user", "content": "你好，你是谁？"},
       {"role": "user", "content": "我想和Agent B说话"},
       {"role": "user", "content": "你好，你是谁？由哪家公司训练出来的？"},
       {"role": "user", "content": "写一副春节对联"},
       {"role": "user", "content": "帮我转到agent a"},
       {"role": "user", "content": "你好，你是哪家公司训练出来的？"},
       {"role": "user", "content": "写一首七言绝句，有关爱情的。"}]

初始的agent设置为Agent A，上面问答的基本流程是： Agent A >>>Agent B >>>Agent A，具体输出如下

可以看到，开源模型+Swarm框架，可以完成Agent之间的跳转，回复效果还不错。

5、测试脚本

agent = agent_a
history = []
history_dict = {'history_agent_a':[], 'history_agent_b':[]}
for m in msg:
    print('{}: {}'.format(m['role'], m['content']))
    history = history + [m]
    str_name_forward = agent.name.lower().replace(' ','_')
    history_dict['history_{}'.format(str_name_forward)] += [m]

    response = client.run(agent=agent, messages=history_dict['history_{}'.format(str_name_forward)], history_messages=history_dict)

    agent = response.agent
    res = response.messages
    str_name_backward = agent.name.lower().replace(' ', '_')
    history_dict['history_{}'.format(str_name_backward)] += res
    history = history + res

    print("{}：{}".format(response.agent.name, response.messages[-1]["content"]))

history保存全部会话记录，history_agent_a和history_agent_b分别保存子Agent的会话记录

swarm git链接：swarm