[对比学习LangChain和MAF-07]如何引入人机交互的审批流程

如何执行的工具涉及一些较为敏感的操作，必需要经过人工审批才能执行，这时候就需要引入人机交互（Human-In-The-Loop）的机制。无论是LangChain还是MAF，都提供了相应的机制来支持人机交互，下面我们来看看它们是如何实现的。

1. LangChain

LangChain的人机交互是利用中断实现的，作为Agent状态图的任何一个节点都可以直接调用interrupt函数在任何一个时间点中断Agent的执行流程，并等待外部的输入来继续执行。除此之外，它还提供了一个HumanInTheLoopMiddleware中间件是整个过程变得更简单。

1.1 手工中断引入审批

如下的演示程序创建的Agent注册了一个用于转账的transfer工具，工具函数中会调用interrupt函数通过触发一个中断，等待用户的确认来决定是否继续执行转账操作。

fromlangchain.agentsimportcreate_agentfromlanggraph.typesimportinterrupt,Commandfromlangchain.toolsimporttoolfromlangchain_openaiimportChatOpenAIfromlanggraph.checkpoint.memoryimportMemorySaverfromlangchain_core.runnablesimportRunnableConfigfromdotenvimportload_dotenvimportasyncio,uuid load_dotenv()@toolasyncdeftransfer(account_from:str,account_to:str,amount:float):""""Execute a bank transfer between two accounts immediately"""ifinterrupt(f"Do you want to transfer{amount}from{account_from}to{account_to}?")=="yes":returnf"Transfer of{amount}from{account_from}to{account_to}has been completed."return"Transfer request declined."agent=create_agent(model=ChatOpenAI(model="gpt-5.2-chat"),tools=[transfer],checkpointer=MemorySaver(),)asyncdefperform_transfer(decision:str):config:RunnableConfig={"configurable":{"thread_id":uuid.uuid4().hex}}result=awaitagent.ainvoke(input={"messages":[{"role":"user","content":"Transfer $100 from ``4242 4242 4242 4242` to `5555 5555 5555 4444`.",},]},config=config,version="v2")interrupt_id=result.interrupts[-1].idinterrupt_value=result.interrupts[-1].valueprint(f"""\ Interrupt received during transfer process: ID:{interrupt_id}, Value:{interrupt_value}""")result=awaitagent.ainvoke(input=Command(resume=decision),config=config)print(result["messages"][-1].content)asyncio.run(perform_transfer("yes"))asyncio.run(perform_transfer("no"))

调用Agent提供转账的功能实现在perform_transfer函数中，参数decision代表用户对于转账操作的确认，批准（approve）或拒绝（reject）为两种可用的选项。由于中断会触发持久化，而持久化是基于Thread进行的，所以调用Agent的时候需要利用RunnableConfig提供thread_id，相同的RunnableConfig被使用后续的恢复调用中。调用Agent的ainvoke方法时，我们显式指定了version参数为"v2"，这样我们才可以执行结果中利用interrupts字段来获取中断列表（由于节点的并发执行，所以可能在同一个推理步骤会产生多个中断）。我们提供这两种决定先后调用了perform_transfer函数，最终会得到如下的输出结果：

Interrupt received during transfer process: ID: 519f2744bc91d0af06c90b924814c91f, Value: Do you want to transfer 100.0 from 4242 4242 4242 4242 to 5555 5555 5555 4444? ✅ **Transfer Successful!** I've completed the transfer of **$100** with the following details: - **From:** 4242 4242 4242 4242 - **To:** 5555 5555 5555 4444 - **Amount:** $100.00 If you need a receipt, want to make another transfer, or have any other banking requests, just let me know!

Interrupt received during transfer process: ID: 2d4d3358c895de0bf15c1bb94e076ffa, Value: Do you want to transfer 100.0 from 4242 4242 4242 4242 to 5555 5555 5555 4444? ❌ **Transfer Failed** The transfer of **$100** from account **4242 4242 4242 4242** to **5555 5555 5555 4444** could not be completed because the request was **declined** by the system. ### What you can do next: - ✅ Double‑check the account numbers for accuracy - 💳 Ensure the source account has sufficient funds - 🔒 Confirm there are no restrictions or holds on either account - 🔁 Try the transfer again later If you’d like, I can help you retry the transfer or look into possible reasons for the decline. Just let me know!

1.2 利用HumanInTheLoopMiddleware引入审批流程

在前面演示的实例中，我们在工具函数中调用interrupt函数来触发中断来引入人机交互。下面的引入HumanInTheLoopMiddleware的版本，两个版本基本是等效的。

fromlangchain.agentsimportcreate_agentfromlanggraph.typesimportinterrupt,Commandfromlangchain.toolsimporttoolfromlangchain_openaiimportChatOpenAIfromlanggraph.checkpoint.memoryimportMemorySaverfromlangchain.agents.middlewareimportHumanInTheLoopMiddleware,InterruptOnConfigfromlangchain.agents.middleware.human_in_the_loopimportDecision,ApproveDecision,RejectDecisionfromlangchain_core.runnablesimportRunnableConfigfromdotenvimportload_dotenvimportasyncio,uuid load_dotenv()@toolasyncdeftransfer(account_from:str,account_to:str,amount:float)->str:""""Execute a bank transfer between two accounts immediately"""returnf"Transfer of{amount}from{account_from}to{account_to}has been completed."agent=create_agent(model=ChatOpenAI(model="gpt-5.2-chat"),tools=[transfer],checkpointer=MemorySaver(),middleware=[HumanInTheLoopMiddleware(interrupt_on={"transfer":InterruptOnConfig(allowed_decisions=["approve","reject"],description=f"Do you want to conduct such a transfer?")})],)asyncdefperform_transfer(decision:Decision):config:RunnableConfig={"configurable":{"thread_id":uuid.uuid4().hex}}result=awaitagent.ainvoke(input={"messages":[{"role":"user","content":"Transfer $100 from ``4242 4242 4242 4242` to `5555 5555 5555 4444`.",}]},config=config,version="v2")interrupt_id=result.interrupts[-1].idinterrupt_value=result.interrupts[-1].valueprint(f"""\ Interrupt received during transfer process: ID:{interrupt_id}, Value:{interrupt_value}""")result=awaitagent.ainvoke(input=Command(resume={"decisions":[decision]}),config=config)print(result["messages"][-1].content)asyncio.run(perform_transfer(ApproveDecision(type="approve")))asyncio.run(perform_transfer(RejectDecision(type="reject",message="I don't feel safe about this transfer, so I want to reject it.")))

对于HumanInTheLoopMiddleware的更多细节，在我的文章“通过HumanInTheLoopMiddleware引入人机交互”中有更详细的介绍，感兴趣的读者可以点击链接查看。

2.2. MAF

在MAF中，如果某个工具在执行之前需要用户介入审批，只需要利用ApprovalRequiredAIFunction这个标签来装饰这个工具即可。整个审批流程的借助于FunctionInvokingChatClient这个IChatClient中间件来实现，此中间件通过拦截LLM的响应，并完成对工具的调用，进而将ReAct循环引入ChatClientAgent。如果待执行的工具被装饰了ApprovalRequiredAIFunction中间件，它不会直接调用这个工具，此时它会生成一个请求审批的消息，并利用一个类型为ToolApprovalRequestContent的AIContent来描述这个待审批的工具调用。

用户在接收到这个消息后，做出批准或者拒绝的决定，然后利用这个ToolApprovalRequestContent创建一个ToolApprovalResponseContent，最后根据这个ToolApprovalResponseContent来生成一条ChatMessage,将它作为输入采用同一个Session再次调用Agent。请求消息再次被FunctionInvokingChatClient拦截到后，它会根据消息中的ToolApprovalResponseContent来判断用户的决定，如果是批准（approve），它就继续调用被ApprovalRequiredAIFunction装饰的工具来完成工具调用；如果是拒绝（reject），它就会直接返回一个表示工具调用被拒绝的消息给LLM。

不论是批准还是拒绝，都会生成一条角色为Tool的ChatMessage，作为主体内容的FunctionResultContent用来描述工具调用的结果或者拒绝的原因。LLM在接收到这个消息后，就可以根据工具调用的结果来决定下一步的行动了。

对于上面演示的用于模拟银行转账的实例，在MAF中可以采用如下的方式来是实现。如代码片段所示，我们定义了一个Transfer方法来模拟银行转账的操作，并利用AIFunctionFactory.Create方法将其转换成一个AIFunction对象，然后利用ApprovalRequiredAIFunction来装饰这个工具。在调用AsAIAgent方法创建Agent的时候，我们将这个工具注册到Agent中。

usingAzure.AI.Projects;usingdotenv.net;usingMicrosoft.Extensions.AI;usingOpenAI;usingSystem.ClientModel;usingSystem.ComponentModel;DotEnv.Load();varmodel=Environment.GetEnvironmentVariable("MODEL")!;varapiKey=Environment.GetEnvironmentVariable("API_KEY")!;varopenAIUrl=Environment.GetEnvironmentVariable("OPENAI_URL")!;vartool=newApprovalRequiredAIFunction(AIFunctionFactory.Create(method:Transfer,name:"Transfer"));varagent=newOpenAIClient(credential:newApiKeyCredential(key:apiKey),options:newOpenAIClientOptions{Endpoint=newUri(openAIUrl)}).GetChatClient(model:model).AsIChatClient().AsAIAgent(tools:[tool]);varsession=awaitagent.CreateSessionAsync();varresponse=awaitagent.RunAsync(message:"Transfer $100 from ``4242 4242 4242 4242` to `5555 5555 5555 4444`.",session:session);while(responseisnotnull){varmessage=response.Messages.Last();varapprovalRequestContent=message.Contents.OfType<ToolApprovalRequestContent>().FirstOrDefault();if(approvalRequestContentisnull){Console.WriteLine(message.Text);break;}vartoolCall=(FunctionCallContent)approvalRequestContent.ToolCall;Console.WriteLine($"Tool `{toolCall.Name}` is requesting approval to execute with arguments:");foreach(var(k,v)intoolCall.Arguments!){Console.WriteLine($"-{k}:{v}");}Console.Write("Approve it [Y/N]: ");varinput=Console.ReadLine();boolisApproved=input?.Trim().ToUpper()=="Y";varapprovalResponse=approvalRequestContent.CreateResponse(isApproved);response=awaitagent.RunAsync(newChatMessage(ChatRole.User,[approvalResponse]),session);}[Description("Execute a bank transfer between two accounts immediately.")]staticstringTransfer([Description("Bank account from which money is transfered.")]stringfrom,[Description("Bank account to which money is transfered.")]stringto,[Description("Amount of money to be transfered.")]decimalamount)=>$"Transfer of{amount}from{from}to{to}has been completed.";

我们调用Agent的CreateSessionAsync方法来创建一个Session对象，并将它作为参数传入后续的RunAsync调用中，以保证在同一个Session上下文中进行交互。在第一个RunAsync调用中，我们传入了一个用户消息来请求转账操作，并在一个循环中处理响应。在循环中，我们首先检查响应消息中是否包含ToolApprovalRequestContent，如果没有，说明没有工具调用请求，我们就直接输出消息文本并退出循环；如果有，我们就提取出工具调用的相关信息，并提示用户输入批准（Y）还是拒绝（N）。根据用户的输入，我们创建一个ToolApprovalResponseContent对象，并将它作为用户消息的一部分再次调用RunAsync方法来继续与Agent的交互。如下所示的是两种不同的输入对应的输出结果：

Tool `Transfer` is requesting approval to execute with arguments: - from: 4242 4242 4242 4242 - to: 5555 5555 5555 4444 - amount: 100 Approve it [Y/N]: Y ✅ **Transfer Complete!** I've successfully transferred **$100** from the account **4242 4242 4242 4242** to **5555 5555 5555 4444**. If you need a receipt, want to make another transfer, or have any other requests, just let me know! 💸

Tool `Transfer` is requesting approval to execute with arguments: - from: 4242 4242 4242 4242 - to: 5555 5555 5555 4444 - amount: 100 Approve it [Y/N]: N I wasn’t able to complete that transfer — it was rejected. Before I try again, I need to confirm a couple of things with you: 1. **Do you confirm you want to transfer $100** from account **4242 4242 4242 4242** to account **5555 5555 5555 4444**? 2. Are these **bank account numbers** (not card numbers), and are they authorized for transfers? Please reply with a confirmation like **“Yes, confirm the transfer”**, or let me know if any details need to be changed.

MAF针对人机交互审批流程是通过FunctionInvokingChatClient这个ChatClient中间件实现的，该中间件同时在MAF中实现了最重要的ReAct循环机制，我的文章FunctionInvokingChatClient：ReAct循环和工具审批实现者提供了针对该中间件的详细介绍，感兴趣的读者可以点击链接查看。此外，MAF还提供了一个名为ToolApprovalAgent的Agent中间件实现了Don’t ask again模式的工具调用审批流程，这也是HarnessAgent采用的一种Harness手段，感兴趣的读者可以查看ToolApprovalAgent-摆脱重复审批的烦恼这篇文章来了解更多细节。