langchain-middleware

面向生产环境的 LangChain Agent 中间件模式：

• HumanInTheLoopMiddleware / humanInTheLoopMiddleware：在执行危险的工具调用前暂停，以供人工审批
• 自定义中间件：拦截工具调用，用于错误处理、日志记录和重试逻辑
• Command 恢复执行：根据人工决策（批准、编辑、拒绝）继续执行

要求： 所有 HITL 工作流都需要 Checkpointer 和 thread_id 配置。

---

人在回路中

设置一个带有 HITL 中间件的 Agent，在发送邮件前暂停以供审批。

from langchain.agents import create_agent
from langchain.agents.middleware import HumanInTheLoopMiddleware
from langgraph.checkpoint.memory import MemorySaver
from langchain.tools import tool

@tool
def send_email(to: str, subject: str, body: str) -> str:
    """Send an email."""
    return f"Email sent to {to}"

agent = create_agent(
    model="gpt-4.1",
    tools=[send_email],
    checkpointer=MemorySaver(),  # Required for HITL
    middleware=[
        HumanInTheLoopMiddleware(
            interrupt_on={
                "send_email": {"allowed_decisions": ["approve", "edit", "reject"]},
            }
        )
    ],
)

设置一个带有 HITL 的 Agent，在发送邮件前暂停以供人工审批。

import { createAgent, humanInTheLoopMiddleware } from "langchain";
import { MemorySaver } from "@langchain/langgraph";
import { tool } from "@langchain/core/tools";
import { z } from "zod";

const sendEmail = tool(
  async ({ to, subject, body }) => `Email sent to ${to}`,
  {
    name: "send_email",
    description: "Send an email",
    schema: z.object({ to: z.string(), subject: z.string(), body: z.string() }),
  }
);

const agent = createAgent({
  model: "anthropic:claude-sonnet-4-5",
  tools: [sendEmail],
  checkpointer: new MemorySaver(),
  middleware: [
    humanInTheLoopMiddleware({
      interruptOn: { send_email: { allowedDecisions: ["approve", "edit", "reject"] } },
    }),
  ],
});

运行 Agent，检测中断，然后在人工批准后继续执行。

from langgraph.types import Command

config = {"configurable": {"thread_id": "session-1"}}

# Step 1: Agent runs until it needs to call tool
result1 = agent.invoke({
    "messages": [{"role": "user", "content": "Send email to [email protected]"}]
}, config=config)

# Check for interrupt
if "__interrupt__" in result1:
    print(f"Waiting for approval: {result1['__interrupt__']}")

# Step 2: Human approves
result2 = agent.invoke(
    Command(resume={"decisions": [{"type": "approve"}]}),
    config=config
)

运行 Agent，检测中断，然后在人工批准后继续执行。

import { Command } from "@langchain/langgraph";

const config = { configurable: { thread_id: "session-1" } };

// Step 1: Agent runs until it needs to call tool
const result1 = await agent.invoke({
  messages: [{ role: "user", content: "Send email to [email protected]" }]
}, config);

// Check for interrupt
if (result1.__interrupt__) {
  console.log(`Waiting for approval: ${result1.__interrupt__}`);
}

// Step 2: Human approves
const result2 = await agent.invoke(
  new Command({ resume: { decisions: [{ type: "approve" }] } }),
  config
);

如果原始值需要更正，请在批准前编辑工具参数。

# Human edits the arguments — edited_action must include name + args
result2 = agent.invoke(
    Command(resume={
        "decisions": [{
            "type": "edit",
            "edited_action": {
                "name": "send_email",
                "args": {
                    "to": "[email protected]",  # Fixed email
                    "subject": "Project Meeting - Updated",
                    "body": "...",
                },
            },
        }]
    }),
    config=config
)

如果原始值需要更正，请在批准前编辑工具参数。

// Human edits the arguments — editedAction must include name + args
const result2 = await agent.invoke(
  new Command({
    resume: {
      decisions: [{
        type: "edit",
        editedAction: {
          name: "send_email",
          args: {
            to: "[email protected]",  // Fixed email
            subject: "Project Meeting - Updated",
            body: "...",
          },
        },
      }]
    }
  }),
  config
);

拒绝工具调用，并提供说明拒绝原因的反馈。

# Human rejects
result2 = agent.invoke(
    Command(resume={
        "decisions": [{
            "type": "reject",
            "feedback": "Cannot delete customer data without manager approval",
        }]
    }),
    config=config
)

根据风险级别，为每个工具配置不同的 HITL 策略。

agent = create_agent(
    model="gpt-4.1",
    tools=[send_email, read_email, delete_email],
    checkpointer=MemorySaver(),
    middleware=[
        HumanInTheLoopMiddleware(
            interrupt_on={
                "send_email": {"allowed_decisions": ["approve", "edit", "reject"]},
                "delete_email": {"allowed_decisions": ["approve", "reject"]},  # No edit
                "read_email": False,  # No HITL for reading
            }
        )
    ],
)

### 可配置项

• 哪些工具需要审批（按工具设置策略）
• 每个工具允许的决策（批准、编辑、拒绝）
• 自定义中间件钩子：before_model、after_model、wrap_tool_call、before_agent、after_agent
• 特定工具的中间件（仅应用于某些工具）

---

自定义中间件钩子

提供六种装饰器钩子。有两种模式：

• 包装钩子（wrap_tool_call、wrap_model_call）：(request, handler) — 调用 handler(request) 继续执行，或提前返回以短路执行。
• 前置/后置钩子（before_model、after_model、before_agent、after_agent）：(state, runtime) — 检查或修改状态。返回 None 或包含状态更新的字典。

`@wrap_tool_call` 会拦截工具执行。**不要使用 `yield`** — 这会创建一个生成器并导致 `NotImplementedError`。

from langchain.agents.middleware import wrap_tool_call

@wrap_tool_call
def retry_middleware(request, handler):
    for attempt in range(3):
        try:
            return handler(request)
        except Exception:
            if attempt == 2:
                raise

@wrap_tool_call
def guard_middleware(request, handler):
    if request.tool_call["name"] == "dangerous_tool":
        return "This tool is disabled"  # short-circuit
    return handler(request)

`createMiddleware({ wrapToolCall })` 会拦截工具执行。

import { createMiddleware } from "langchain";

const retryMiddleware = createMiddleware({
  wrapToolCall: async (request, handler) => {
    for (let attempt = 0; attempt < 3; attempt++) {
      try { return await handler(request); }
      catch (e) { if (attempt === 2) throw e; }
    }
  },
});

`before_model` / `after_model` / `before_agent` / `after_agent` 都使用 `(state, runtime)` 签名。

from langchain.agents.middleware import before_model, after_model

@before_model
def log_calls(state, runtime):
    print(f"Calling model with {len(state['messages'])} messages")

@after_model
def check_output(state, runtime):
    print(f"Model responded")

通过 `createMiddleware` 创建的所有前置/后置钩子都使用相同的 `(state, runtime)` 签名。

import { createMiddleware } from "langchain";

const loggingMiddleware = createMiddleware({
  beforeModel: (state, runtime) => {
    console.log(`Calling model with ${state.messages.length} messages`);
  },
  afterModel: (state, runtime) => {
    console.log("Model responded");
  },
});

### 不可配置项

• 工具执行后中断（必须在执行前中断）
• 跳过 HITL 所需的 checkpointer

HITL 中间件需要 checkpointer 来持久化状态。

# WRONG
agent = create_agent(model="gpt-4.1", tools=[send_email], middleware=[HumanInTheLoopMiddleware({...})])

# CORRECT
agent = create_agent(
    model="gpt-4.1", tools=[send_email],
    checkpointer=MemorySaver(),  # Required
    middleware=[HumanInTheLoopMiddleware({...})]
)

HITL 需要 checkpointer 来持久化状态。

// WRONG: No checkpointer
const agent = createAgent({
  model: "anthropic:claude-sonnet-4-5", tools: [sendEmail],
  middleware: [humanInTheLoopMiddleware({ interruptOn: { send_email: true } })],
});

// CORRECT: Add checkpointer
const agent = createAgent({
  model: "anthropic:claude-sonnet-4-5", tools: [sendEmail],
  checkpointer: new MemorySaver(),
  middleware: [humanInTheLoopMiddleware({ interruptOn: { send_email: true } })],
});

使用 HITL 时，请始终提供 thread_id 以跟踪对话状态。

# WRONG
agent.invoke(input)  # No config!

# CORRECT
agent.invoke(input, config={"configurable": {"thread_id": "user-123"}})

使用 Command 类在中断后继续执行。

# WRONG
agent.invoke({"resume": {"decisions": [...]}})

# CORRECT
from langgraph.types import Command
agent.invoke(Command(resume={"decisions": [{"type": "approve"}]}), config=config)

使用 Command 类在中断后继续执行。

// WRONG
await agent.invoke({ resume: { decisions: [...] } });

// CORRECT
import { Command } from "@langchain/langgraph";
await agent.invoke(new Command({ resume: { decisions: [{ type: "approve" }] } }), config);