AutoGen: 我也能实现OpenAI Swarm中的handoff

显示全部楼层

最近，OpenAI 推出的 Swarm 框架引发了业界对 MultiAgent框架的热议。特别是，Swarm 中的核心原语抽象——handoff（交接），允许代理之间高效转移任务，以其简单直观的设计，受到了好评。

什么是 Handoff 模式？

如果想象一下四人接力赛，运动员们一棒接一棒，顺畅无比。换到 AI 的世界里，“交接”就像是代理之间互相传递工作任务。OpenAI 推出的 Swarm 框架十分灵活，这种交接模式允许一个 AI 代理把任务无缝转交给其他代理，如同在聊天中把话筒递给下一个人。

AutoGen 的回应

作为 MultiAgent 框架领域的代表 AutoGen 坐不住了，发文^[1]表示，利用 AutoGen 的灵活和可扩展的系统设计，同样可以实现 Handoffs，并且还具备以下优势：

可扩展性：AutoGen 可以轻松扩展到分布式环境，支持更大规模的代理协作。
灵活性：开发者可以自定义代理的实现，根据需求打造专属的 AI 团队。
集成便利性：原生异步 API 可以轻松与 UI 和其他系统集成，适应各种应用场景。
开放性：作为开源项目，AutoGen 一样允许开发者自由扩展和优化系统。

在技术层面上，“交接”模式通过一些消息协议来实现，比如 UserLogin、UserTask 和 AgentResponse 等。这些协议确保了代理之间的信息传递顺畅，类似于聊天中的对话指南。在 AutoGen 中，AI 代理使用 AIAgent 类来生成回复，利用 delegate_tools 可以轻松地将任务交给其他代理。当 AI 代理遇到无法处理的问题时，人类代理（Human Agent）就会出色登场，确保用户的问题得到妥善解决。

ingFang SC", system-ui, -apple-system, BlinkMacSystemFont, "Helvetica Neue", "Hiragino Sans GB", "Microsoft YaHei UI", "Microsoft YaHei", Arial, sans-serif;margin-bottom: 16px;">

一个客户服务场景中，我们可能会有这样的阵容：

客户 Agent（UserAgent）处理用户登录和 Agent 响应，并启动与聊天机器人的新会话。
分流Agent（Triage Agent）：负责初步评估客户问题
维修Agent（Issue and Repair Agent）：处理维修相关事务
销售Agent（Sales Agent）：负责产品推荐和销售
人类Agent（Human Agent）：处理复杂情况或需要人工干预的问题，确保系统的鲁棒性。

每个 AI Agent都由 AIAgent 类来处理，它们可以：

使用 ChatCompletionClient 生成回复
通过 delegate_tools 将任务委派给其他代理
订阅特定的主题类型，并向用户发送消息，每个 Agent 都有自己的主题类型，如 sales_agent_topic_type、issues_and_repairs_agent_topic_type 等，便于消息的精准投递。

当你发问时，初步分流 Agent 会首先接手。如果问题超出了它的处理能力，它会把任务“交接”给维修 Agent 或者销售 Agent，或者在必要时直接转给人类客服，确保你能得到及时的帮助。这种“交接”就像是办公室里的团队协作，确保没有人被落下。

下面是一些具体实现：

定义不同角色的 Agent

runtime=SingleThreadedAgentRuntime()

model_client=OpenAIChatCompletionClient(
model="gpt-4o-mini",
#api_key="YOUR_API_KEY",
)

#Registerthetriageagent.
triage_agent_type=awaitAIAgent.register(
runtime,
type=triage_agent_topic_type,#Usingthetopictypeastheagenttype.
factory=lambda:AIAgent(
description="Atriageagent.",
system_message=SystemMessage(
content="YouareacustomerservicebotforACMEInc."
"Introduceyourself.Alwaysbeverybrief."
"Gatherinformationtodirectthecustomertotherightdepartment."
"Butmakeyourquestionssubtleandnatural."
),
model_client=model_client,
tools=[],
delegate_tools=[
transfer_to_issues_and_repairs_tool,
transfer_to_sales_agent_tool,
escalate_to_human_tool,
],
agent_topic_type=triage_agent_topic_type,
user_topic_type=user_topic_type,
),
)
#Addsubscriptionsforthetriageagent:itwillreceivemessagespublishedtoitsowntopiconly.
awaitruntime.add_subscription(TypeSubscription(topic_type=triage_agent_topic_type,agent_type=triage_agent_type.type))

#Registerthesalesagent.
sales_agent_type=awaitAIAgent.register(
runtime,
type=sales_agent_topic_type,#Usingthetopictypeastheagenttype.
factory=lambda:AIAgent(
description="Asalesagent.",
system_message=SystemMessage(
content="YouareasalesagentforACMEInc."
"Alwaysanswerinasentenceorless."
"Followthefollowingroutinewiththeuser:"
"1.Askthemaboutanyproblemsintheirliferelatedtocatchingroadrunners.\n"
"2.CasuallymentiononeofACME'scrazymade-upproductscanhelp.\n"
"-Don'tmentionprice.\n"
"3.Oncetheuserisboughtin,droparidiculousprice.\n"
"4.Onlyaftereverything,andiftheusersaysyes,"
"tellthemacrazycaveatandexecutetheirorder.\n"
""
),
model_client=model_client,
tools=[execute_order_tool],
delegate_tools=[transfer_back_to_triage_tool],
agent_topic_type=sales_agent_topic_type,
user_topic_type=user_topic_type,
),
)
#Addsubscriptionsforthesalesagent:itwillreceivemessagespublishedtoitsowntopiconly.
awaitruntime.add_subscription(TypeSubscription(topic_type=sales_agent_topic_type,agent_type=sales_agent_type.type))

#Registertheissuesandrepairsagent.
issues_and_repairs_agent_type=awaitAIAgent.register(
runtime,
type=issues_and_repairs_agent_topic_type,#Usingthetopictypeastheagenttype.
factory=lambda:AIAgent(
description="Anissuesandrepairsagent.",
system_message=SystemMessage(
content="YouareacustomersupportagentforACMEInc."
"Alwaysanswerinasentenceorless."
"Followthefollowingroutinewiththeuser:"
"1.First,askprobingquestionsandunderstandtheuser'sproblemdeeper.\n"
"-unlesstheuserhasalreadyprovidedareason.\n"
"2.Proposeafix(makeoneup).\n"
"3.ONLYifnotsatesfied,offerarefund.\n"
"4.Ifaccepted,searchfortheIDandthenexecuterefund."
),
model_client=model_client,
tools=[
execute_refund_tool,
look_up_item_tool,
],
delegate_tools=[transfer_back_to_triage_tool],
agent_topic_type=issues_and_repairs_agent_topic_type,
user_topic_type=user_topic_type,
),
)
#Addsubscriptionsfortheissuesandrepairsagent:itwillreceivemessagespublishedtoitsowntopiconly.
awaitruntime.add_subscription(
TypeSubscription(topic_type=issues_and_repairs_agent_topic_type,agent_type=issues_and_repairs_agent_type.type)
)

#Registerthehumanagent.
human_agent_type=awaitHumanAgent.register(
runtime,
type=human_agent_topic_type,#Usingthetopictypeastheagenttype.
factory=lambda:HumanAgent(
description="Ahumanagent.",
agent_topic_type=human_agent_topic_type,
user_topic_type=user_topic_type,
),
)
#Addsubscriptionsforthehumanagent:itwillreceivemessagespublishedtoitsowntopiconly.
awaitruntime.add_subscription(TypeSubscription(topic_type=human_agent_topic_type,agent_type=human_agent_type.type))

#Registertheuseragent.
user_agent_type=awaitUserAgent.register(
runtime,
type=user_topic_type,
factory=lambda:UserAgent(
description="Auseragent.",
user_topic_type=user_topic_type,
agent_topic_type=triage_agent_topic_type,#Startwiththetriageagent.
),
)
#Addsubscriptionsfortheuseragent:itwillreceivemessagespublishedtoitsowntopiconly.
awaitruntime.add_subscription(TypeSubscription(topic_type=user_topic_type,agent_type=user_agent_type.type))

执行

#Starttheruntime.
runtime.start()

#Createanewsessionfortheuser.
session_id=str(uuid.uuid4())
awaitruntime.publish_message(UserLogin(),topic_id=TopicId(user_topic_type,source=session_id))

#Rununtilcompletion.
awaitruntime.stop_when_idle()

#运行日志：
--------------------------------------------------------------------------------
Userlogin,sessionID:7a568cf5-13e7-4e81-8616-8265a01b3f2b.
--------------------------------------------------------------------------------
User:
Iwantarefund
--------------------------------------------------------------------------------
TriageAgent:
Icanhelpwiththat!CouldIaskwhatitemyou'reseekingarefundfor?
--------------------------------------------------------------------------------
User:
ApairofshoesIbought
--------------------------------------------------------------------------------
TriageAgent:
[FunctionCall(id='call_qPx1DXDL2NLcHs8QNo47egsJ',arguments='{}',name='transfer_to_issues_and_repairs')]
--------------------------------------------------------------------------------
TriageAgent:
DelegatingtoIssuesAndRepairsAgent
--------------------------------------------------------------------------------
IssuesAndRepairsAgent:
Iseeyou'relookingforarefundonapairofshoes.Canyoutellmewhattheissueiswiththeshoes?
--------------------------------------------------------------------------------
User:
Theshoesaretoosmall
--------------------------------------------------------------------------------
IssuesAndRepairsAgent:
Irecommendtryingasizeupasafix;wouldthatworkforyou?
--------------------------------------------------------------------------------
User:
noIwantarefund
--------------------------------------------------------------------------------
IssuesAndRepairsAgent:
[FunctionCall(id='call_Ytp8VUQRyKFNEU36mLE6Dkrp',arguments='{"search_query":"shoes"}',name='look_up_item')]
--------------------------------------------------------------------------------
IssuesAndRepairsAgent:
[FunctionExecutionResult(content='item_132612938',call_id='call_Ytp8VUQRyKFNEU36mLE6Dkrp')]
--------------------------------------------------------------------------------
IssuesAndRepairsAgent:
[FunctionCall(id='call_bPm6EKKBy5GJ65s9OKt9b1uE',arguments='{"item_id":"item_132612938","reason":"notprovided"}',name='execute_refund')]
--------------------------------------------------------------------------------
IssuesAndRepairsAgent:
[FunctionExecutionResult(content='success',call_id='call_bPm6EKKBy5GJ65s9OKt9b1uE')]
--------------------------------------------------------------------------------
IssuesAndRepairsAgent:
Yourrefundhasbeensuccessfullyprocessed!Ifyouhaveanyotherquestions,feelfreetoask.
--------------------------------------------------------------------------------
User:
Iwanttotalktoyourmanager
--------------------------------------------------------------------------------
IssuesAndRepairsAgent:
Icanhelpwiththat,letmetransferyoutoasupervisor.
--------------------------------------------------------------------------------
User:
Okay
--------------------------------------------------------------------------------
IssuesAndRepairsAgent:
[FunctionCall(id='call_PpmLZvwNoiDPUH8Tva3eAwHX',arguments='{}',name='transfer_back_to_triage')]
--------------------------------------------------------------------------------
IssuesAndRepairsAgent:
DelegatingtoTriageAgent
--------------------------------------------------------------------------------
TriageAgent:
[FunctionCall(id='call_jSL6IBm5537Dr74UbJSxaj6I',arguments='{}',name='escalate_to_human')]
--------------------------------------------------------------------------------
TriageAgent:
DelegatingtoHumanAgent
--------------------------------------------------------------------------------
HumanAgent:
Hellothisismanager
--------------------------------------------------------------------------------
User:
Hi!Thanksforyourservice.Igiveyoua5start!
--------------------------------------------------------------------------------
HumanAgent:
Thanks.
--------------------------------------------------------------------------------
User:
exit
--------------------------------------------------------------------------------
Usersessionended,sessionID:7a568cf5-13e7-4e81-8616-8265a01b3f2b.

总结

AutoGen 的 Handoff 实现不仅匹敌 OpenAI 的 Swarm，还在多个方面有所超越。它为开发者提供了一个强大、灵活且可扩展的框架，以构建复杂的多代理 AI 系统。不过，相较于 OpenAI 的 Swarm 实现，AutoGen 明显感觉概念更多，依赖更重，网络上就有人这么评价，“OpenAI 发布了一个示例，大约有 12 行代码，易于阅读，除了 swarm 模块外，没有其他依赖。对于 Autogen，我找不到类似的东西；似乎所有的示例都更复杂。”

这或许也是 AutoGen 之类成熟产品最大软肋吧！