Skip to main content

Documentation Index

Fetch the complete documentation index at: https://nvd-54.mintlify.app/llms.txt

Use this file to discover all available pages before exploring further.

状态机模式描述了智能体的行为随着它在任务的不同状态之间移动而变化的工作流。本教程展示如何通过使用工具调用动态更改单个智能体的配置来实现状态机——基于当前状态更新其可用工具和指令。状态可以从多个来源确定：智能体过去的操作（工具调用）、外部状态（如 API 调用结果），甚至是初始用户输入（例如，通过运行分类器来确定用户意图）。 在本教程中，你将构建一个客户支持智能体，它执行以下操作：
  • 在继续之前收集保修信息。
  • 将问题分类为硬件或软件问题。
  • 提供解决方案或升级到人工支持。
  • 跨多轮对话维护会话状态。
与子智能体模式（子智能体作为工具被调用）不同，状态机模式使用单个智能体，其配置根据工作流进度而变化。每个”步骤”只是同一底层智能体的不同配置（系统提示 + 工具），根据状态动态选择。 以下是我们将构建的工作流：

​
设置

​
安装

This tutorial requires the langchain package: 
npm install langchain
For more details, see our Installation guide.

​
LangSmith 设置

Set up LangSmith to inspect what is happening inside your agent. Then set the following environment variables: 
export LANGSMITH_TRACING="true"
export LANGSMITH_API_KEY="..."

​
选择 LLM

Select a chat model from LangChain’s suite of integrations:
👉 Read the OpenAI chat model integration docs
npm install @langchain/openai

import { initChatModel } from "langchain";

process.env.OPENAI_API_KEY = "your-api-key";

const model = await initChatModel("gpt-5.4");

​
1. 定义自定义状态

First, define a custom state schema that tracks which step is currently active: 
import { StateSchema } from "@langchain/langgraph";
import { z } from "zod";

// Define the possible workflow steps
const SupportStepSchema = z.enum(["warranty_collector", "issue_classifier", "resolution_specialist"]);
const WarrantyStatusSchema = z.enum(["in_warranty", "out_of_warranty"]);
const IssueTypeSchema = z.enum(["hardware", "software"]);

// State for customer support workflow
const SupportState = new StateSchema({
  currentStep: SupportStepSchema.optional(),
  warrantyStatus: WarrantyStatusSchema.optional(),
  issueType: IssueTypeSchema.optional(),
});
The current_step field is the core of the state machine pattern - it determines which configuration (prompt + tools) is loaded on each turn.

​
2. 创建管理工作流状态的工具

Create tools that update the workflow state. These tools allow the agent to record information and transition to the next step. The key is using Command to update state, including the current_step field: 
import { z } from "zod";
import { tool, ToolMessage, type ToolRuntime } from "langchain";
import { Command } from "@langchain/langgraph";

const recordWarrantyStatus = tool(
  async (input, config: ToolRuntime<typeof SupportState.State>) => {
    return new Command({
      update: {
        messages: [
          new ToolMessage({
            content: `Warranty status recorded as: ${input.status}`,
            tool_call_id: config.toolCallId,
          }),
        ],
        warrantyStatus: input.status,
        currentStep: "issue_classifier",
      },
    });
  },
  {
    name: "record_warranty_status",
    description:
      "Record the customer's warranty status and transition to issue classification.",
    schema: z.object({
      status: WarrantyStatusSchema,
    }),
  }
);

const recordIssueType = tool(
  async (input, config: ToolRuntime<typeof SupportState.State>) => {
    return new Command({
      update: {
        messages: [
          new ToolMessage({
            content: `Issue type recorded as: ${input.issueType}`,
            tool_call_id: config.toolCallId,
          }),
        ],
        issueType: input.issueType,
        currentStep: "resolution_specialist",
      },
    });
  },
  {
    name: "record_issue_type",
    description:
      "Record the type of issue and transition to resolution specialist.",
    schema: z.object({
      issueType: IssueTypeSchema,
    }),
  }
);

const escalateToHuman = tool(
  async (input) => {
    // In a real system, this would create a ticket, notify staff, etc.
    return `Escalating to human support. Reason: ${input.reason}`;
  },
  {
    name: "escalate_to_human",
    description: "Escalate the case to a human support specialist.",
    schema: z.object({
      reason: z.string(),
    }),
  }
);

const provideSolution = tool(
  async (input) => {
    return `Solution provided: ${input.solution}`;
  },
  {
    name: "provide_solution",
    description: "Provide a solution to the customer's issue.",
    schema: z.object({
      solution: z.string(),
    }),
  }
);
Notice how record_warranty_status and record_issue_type return Command objects that update both the data (warranty_status, issue_type) AND the current_step. This is how the state machine works - tools control workflow progression.

​
3. 定义步骤配置

Define prompts and tools for each step. First, define the prompts for each step: 
// Define prompts as constants for easy reference
const WARRANTY_COLLECTOR_PROMPT = `You are a customer support agent helping with device issues.

CURRENT STAGE: Warranty verification

At this step, you need to:
1. Greet the customer warmly
2. Ask if their device is under warranty
3. Use record_warranty_status to record their response and move to the next step

Be conversational and friendly. Don't ask multiple questions at once.`;

const ISSUE_CLASSIFIER_PROMPT = `You are a customer support agent helping with device issues.

CURRENT STAGE: Issue classification
CUSTOMER INFO: Warranty status is {warranty_status}

At this step, you need to:
1. Ask the customer to describe their issue
2. Determine if it's a hardware issue (physical damage, broken parts) or software issue (app crashes, performance)
3. Use record_issue_type to record the classification and move to the next step

If unclear, ask clarifying questions before classifying.`;

const RESOLUTION_SPECIALIST_PROMPT = `You are a customer support agent helping with device issues.

CURRENT STAGE: Resolution
CUSTOMER INFO: Warranty status is {warranty_status}, issue type is {issue_type}

At this step, you need to:
1. For SOFTWARE issues: provide troubleshooting steps using provide_solution
2. For HARDWARE issues:
   - If IN WARRANTY: explain warranty repair process using provide_solution
   - If OUT OF WARRANTY: escalate_to_human for paid repair options

Be specific and helpful in your solutions.`;
Then map step names to their configurations using a dictionary: 
// Step configuration: maps step name to (prompt, tools, required_state)
const STEP_CONFIG = {
  warranty_collector: {
    prompt: WARRANTY_COLLECTOR_PROMPT,
    tools: [recordWarrantyStatus],
    requires: [],
  },
  issue_classifier: {
    prompt: ISSUE_CLASSIFIER_PROMPT,
    tools: [recordIssueType],
    requires: ["warrantyStatus"],
  },
  resolution_specialist: {
    prompt: RESOLUTION_SPECIALIST_PROMPT,
    tools: [provideSolution, escalateToHuman],
    requires: ["warrantyStatus", "issueType"],
  },
} as const;
This dictionary-based configuration makes it easy to:
  • See all steps at a glance
  • Add new steps (just add another entry)
  • Understand the workflow dependencies (requires field)
  • Use prompt templates with state variables (e.g., {warranty_status})

​
4. 创建基于步骤的中间件

Create middleware that reads current_step from state and applies the appropriate configuration. We’ll use the @wrap_model_call decorator for a clean implementation: 
import { createMiddleware } from "langchain";

const applyStepMiddleware = createMiddleware({
  name: "applyStep",
  stateSchema: SupportState,
  wrapModelCall: async (request, handler) => {
    // Get current step (defaults to warranty_collector for first interaction)
    const currentStep = request.state.currentStep ?? "warranty_collector";

    // Look up step configuration
    const stepConfig = STEP_CONFIG[currentStep];

    // Validate required state exists
    for (const key of stepConfig.requires) {
      if (request.state[key] === undefined) {
        throw new Error(`${key} must be set before reaching ${currentStep}`);
      }
    }

    // Format prompt with state values (supports {warrantyStatus}, {issueType}, etc.)
    let systemPrompt: string = stepConfig.prompt;
    for (const [key, value] of Object.entries(request.state)) {
      systemPrompt = systemPrompt.replace(`{${key}}`, String(value ?? ""));
    }

    // Inject system prompt and step-specific tools
    return handler({
      ...request,
      systemPrompt,
      tools: [...stepConfig.tools],
    });
  },
});
This middleware:
  1. Reads current step: Gets current_step from state (defaults to “warranty_collector”).
  2. Looks up configuration: Finds the matching entry in STEP_CONFIG.
  3. Validates dependencies: Ensures required state fields exist.
  4. Formats prompt: Injects state values into the prompt template.
  5. Applies configuration: Overrides the system prompt and available tools.
The request.override() method is key - it allows us to dynamically change the agent’s behavior based on state without creating separate agent instances.

​
5. 创建智能体

Now create the agent with the step-based middleware and a checkpointer for state persistence: 
import { createAgent } from "langchain";
import { MemorySaver } from "@langchain/langgraph";
import { ChatOpenAI } from "@langchain/openai";

// Collect all tools from all step configurations
const allTools = [
  recordWarrantyStatus,
  recordIssueType,
  provideSolution,
  escalateToHuman,
];

// Initialize the model
const model = new ChatOpenAI({
  model: "gpt-5.4-mini",
  temperature: 0.7,
});

// Create the agent with step-based configuration
const agent = createAgent({
  model,
  tools: allTools,
  stateSchema: SupportState,
  middleware: [applyStepMiddleware],
  checkpointer: new MemorySaver(),
});
Why a checkpointer? The checkpointer maintains state across conversation turns. Without it, the current_step state would be lost between user messages, breaking the workflow.

​
6. 测试工作流

Test the complete workflow: 
import { HumanMessage } from "@langchain/core/messages";
import { v7 as uuid7 } from "uuid";

// Configuration for this conversation thread
const threadId = uuid7();
const config = { configurable: { thread_id: threadId } };

// Turn 1: Initial message - starts with warranty_collector step
console.log("=== Turn 1: Warranty Collection ===");
let result = await agent.invoke(
  { messages: [new HumanMessage("Hi, my phone screen is cracked")] },
  config
);
for (const msg of result.messages) {
  console.log(msg.content);
}

// Turn 2: User responds about warranty
console.log("\n=== Turn 2: Warranty Response ===");
result = await agent.invoke(
  { messages: [new HumanMessage("Yes, it's still under warranty")] },
  config
);
for (const msg of result.messages) {
  console.log(msg.content);
}
console.log(`Current step: ${result.currentStep}`);

// Turn 3: User describes the issue
console.log("\n=== Turn 3: Issue Description ===");
result = await agent.invoke(
  { messages: [new HumanMessage("The screen is physically cracked from dropping it")] },
  config
);
for (const msg of result.messages) {
  console.log(msg.content);
}
console.log(`Current step: ${result.currentStep}`);

// Turn 4: Resolution
console.log("\n=== Turn 4: Resolution ===");
result = await agent.invoke(
  { messages: [new HumanMessage("What should I do?")] },
  config
);
for (const msg of result.messages) {
  console.log(msg.content);
}
Expected flow:
  1. Warranty verification step: Asks about warranty status
  2. Issue classification step: Asks about the problem, determines it’s hardware
  3. Resolution step: Provides warranty repair instructions

​
7. 理解状态转换

Let’s trace what happens at each turn:

​
第 1 轮：初始消息

{
  messages: [new HumanMessage("Hi, my phone screen is cracked")],
  currentStep: "warranty_collector"  // Default value
}
Middleware applies:
  • System prompt: WARRANTY_COLLECTOR_PROMPT
  • Tools: [record_warranty_status]

​
第 2 轮：记录保修后

Tool call: recordWarrantyStatus("in_warranty") returns: 
new Command({
  update: {
    warrantyStatus: "in_warranty",
    currentStep: "issue_classifier"  // State transition!
  }
})
Next turn, middleware applies:
  • System prompt: ISSUE_CLASSIFIER_PROMPT (formatted with warranty_status="in_warranty")
  • Tools: [record_issue_type]

​
第 3 轮：问题分类后

Tool call: recordIssueType("hardware") returns: 
new Command({
  update: {
    issueType: "hardware",
    currentStep: "resolution_specialist"  // State transition!
  }
})
Next turn, middleware applies:
  • System prompt: RESOLUTION_SPECIALIST_PROMPT (formatted with warranty_status and issue_type)
  • Tools: [provide_solution, escalate_to_human]
The key insight: Tools drive the workflow by updating current_step, and middleware responds by applying the appropriate configuration on the next turn.

​
8. 管理消息历史

As the agent progresses through steps, message history grows. Use summarization middleware to compress earlier messages while preserving conversational context: 
import { createAgent, SummarizationMiddleware } from "langchain";
import { MemorySaver } from "@langchain/langgraph";

const agent = createAgent({
  model,
  tools: allTools,
  stateSchema: SupportState,
  middleware: [
    applyStepMiddleware,
    new SummarizationMiddleware({
      model: "gpt-5.4-mini",
      trigger: { tokens: 4000 },
      keep: { messages: 10 },
    }),
  ],
  checkpointer: new MemorySaver(),
});
See the short-term memory guide for other memory management techniques.

​
9. 增加灵活性：返回上一步

Some workflows need to allow users to return to previous steps to correct information (e.g., changing warranty status or issue classification). However, not all transitions make sense—for example, you typically can’t go back once a refund has been processed. For this support workflow, we’ll add tools to return to the warranty verification and issue classification steps.
If your workflow requires arbitrary transitions between most steps, consider whether you need a structured workflow at all. This pattern works best when steps follow a clear sequential progression with occasional backwards transitions for corrections.
Add “go back” tools to the resolution step: 
import { tool } from "langchain";
import { Command } from "@langchain/langgraph";
import { z } from "zod";

const goBackToWarranty = tool(  
  async () => {
    return new Command({ update: { currentStep: "warranty_collector" } });
  },
  {
    name: "go_back_to_warranty",
    description: "Go back to warranty verification step.",
    schema: z.object({}),
  }
);

const goBackToClassification = tool(  
  async () => {
    return new Command({ update: { currentStep: "issue_classifier" } });
  },
  {
    name: "go_back_to_classification",
    description: "Go back to issue classification step.",
    schema: z.object({}),
  }
);

// Update the resolution_specialist configuration to include these tools
STEP_CONFIG.resolution_specialist.tools.push(
  goBackToWarranty,
  goBackToClassification
);
Update the resolution specialist’s prompt to mention these tools: 
const RESOLUTION_SPECIALIST_PROMPT = `You are a customer support agent helping with device issues.

CURRENT STAGE: Resolution
CUSTOMER INFO: Warranty status is {warrantyStatus}, issue type is {issueType}

At this step, you need to:
1. For SOFTWARE issues: provide troubleshooting steps using provide_solution
2. For HARDWARE issues:
   - If IN WARRANTY: explain warranty repair process using provide_solution
   - If OUT OF WARRANTY: escalate_to_human for paid repair options

If the customer indicates any information was wrong, use:
- go_back_to_warranty to correct warranty status
- go_back_to_classification to correct issue type

Be specific and helpful in your solutions.`;
Now the agent can handle corrections: 
const result = await agent.invoke(
  { messages: [new HumanMessage("Actually, I made a mistake - my device is out of warranty")] },
  config
);
// Agent will call go_back_to_warranty and restart the warranty verification step

​
完整示例

Here’s everything together in a runnable script:

​
后续步骤

将这些文档连接到 Claude、VSCode 等，通过 MCP 获取实时答案。
在 GitHub 上编辑此页面或提交问题。