building-ai-agent-on-cloudflare

构建 Cloudflare Agents

你对 Agents SDK 的了解可能已经过时。构建任何 Agent 时，都应优先检索，而非依赖预训练知识。

检索来源

来源	检索方式	适用内容
Agents SDK 文档	https://github.com/cloudflare/agents/tree/main/docs	SDK API、状态、路由、调度
Cloudflare Agents 文档	https://developers.cloudflare.com/agents/	平台集成、部署
Workers 文档	搜索工具或 https://developers.cloudflare.com/workers/	运行时 API、绑定、配置

适用场景

• 用户想构建 AI Agent 或聊天机器人
• 用户需要有状态的实时 AI 交互
• 用户询问 Cloudflare Agents SDK
• 用户想执行定时任务或后台 AI 工作
• 用户需要基于 WebSocket 的 AI 通信

前置条件

• 已启用 Workers 的 Cloudflare 账户
• Node.js 18+ 和 npm/pnpm/yarn
• Wrangler CLI（npm install -g wrangler）

快速开始

npm create cloudflare@latest -- my-agent --template=cloudflare/agents-starter
cd my-agent
npm start

Agent 运行于 http://localhost:8787

核心概念

什么是 Agent？

Agent 是一种有状态、持久运行的 AI 服务，它可以：

• 在请求和重新连接之间保持状态
• 通过 WebSocket 或 HTTP 通信
• 通过 Durable Objects 在 Cloudflare 边缘运行
• 调度任务并调用工具
• 水平扩展（每个用户/会话都有自己的实例）

Agent 生命周期

Client connects → Agent.onConnect() → Agent processes messages
                                    → Agent.onMessage()
                                    → Agent.setState() (persists + syncs)
Client disconnects → State persists → Client reconnects → State restored

Agent 基本结构

import { Agent, Connection } from "agents";

interface Env {
  AI: Ai;  // Workers AI binding
}

interface State {
  messages: Array<{ role: string; content: string }>;
  preferences: Record<string, string>;
}

export class MyAgent extends Agent<Env, State> {
  // Initial state for new instances
  initialState: State = {
    messages: [],
    preferences: {},
  };

  // Called when agent starts or resumes
  async onStart() {
    console.log("Agent started with state:", this.state);
  }

  // Handle WebSocket connections
  async onConnect(connection: Connection) {
    connection.send(JSON.stringify({
      type: "welcome",
      history: this.state.messages,
    }));
  }

  // Handle incoming messages
  async onMessage(connection: Connection, message: string) {
    const data = JSON.parse(message);

    if (data.type === "chat") {
      await this.handleChat(connection, data.content);
    }
  }

  // Handle disconnections
  async onClose(connection: Connection) {
    console.log("Client disconnected");
  }

  // React to state changes
  onStateUpdate(state: State, source: string) {
    console.log("State updated by:", source);
  }

  private async handleChat(connection: Connection, userMessage: string) {
    // Add user message to history
    const messages = [
      ...this.state.messages,
      { role: "user", content: userMessage },
    ];

    // Call AI
    const response = await this.env.AI.run("@cf/meta/llama-3-8b-instruct", {
      messages,
    });

    // Update state (persists and syncs to all clients)
    this.setState({
      ...this.state,
      messages: [
        ...messages,
        { role: "assistant", content: response.response },
      ],
    });

    // Send response
    connection.send(JSON.stringify({
      type: "response",
      content: response.response,
    }));
  }
}

入口点配置

// src/index.ts
import { routeAgentRequest } from "agents";
import { MyAgent } from "./agent";

export default {
  async fetch(request: Request, env: Env) {
    // routeAgentRequest handles routing to /agents/:class/:name
    return (
      (await routeAgentRequest(request, env)) ||
      new Response("Not found", { status: 404 })
    );
  },
};

export { MyAgent };

客户端通过以下地址连接：wss://my-agent.workers.dev/agents/MyAgent/session-id

Wrangler 配置

{
  "name": "my-agent",
  "main": "src/index.ts",
  "compatibility_date": "2024-12-01",
  "ai": { "binding": "AI" },
  "durable_objects": {
    "bindings": [{ "name": "MyAgent", "class_name": "MyAgent" }]
  },
  "migrations": [{ "tag": "v1", "new_sqlite_classes": ["MyAgent"] }]
}

状态管理

读取状态

// Current state is always available
const currentMessages = this.state.messages;
const userPrefs = this.state.preferences;

更新状态

// setState persists AND syncs to all connected clients
this.setState({
  ...this.state,
  messages: [...this.state.messages, newMessage],
});

// Partial updates work too
this.setState({
  preferences: { ...this.state.preferences, theme: "dark" },
});

SQL 存储

对于复杂查询，请使用内嵌的 SQLite 数据库：

// Create tables
await this.sql`
  CREATE TABLE IF NOT EXISTS documents (
    id INTEGER PRIMARY KEY AUTOINCREMENT,
    title TEXT NOT NULL,
    content TEXT,
    created_at DATETIME DEFAULT CURRENT_TIMESTAMP
  )
`;

// Insert
await this.sql`
  INSERT INTO documents (title, content)
  VALUES (${title}, ${content})
`;

// Query
const docs = await this.sql`
  SELECT * FROM documents WHERE title LIKE ${`%${search}%`}
`;

定时任务

Agent 可以调度未来的工作：

async onMessage(connection: Connection, message: string) {
  const data = JSON.parse(message);

  if (data.type === "schedule_reminder") {
    // Schedule task for 1 hour from now
    const { id } = await this.schedule(3600, "sendReminder", {
      message: data.reminderText,
      userId: data.userId,
    });

    connection.send(JSON.stringify({ type: "scheduled", taskId: id }));
  }
}

// Called when scheduled task fires
async sendReminder(data: { message: string; userId: string }) {
  // Send notification, email, etc.
  console.log(`Reminder for ${data.userId}: ${data.message}`);

  // Can also update state
  this.setState({
    ...this.state,
    lastReminder: new Date().toISOString(),
  });
}

调度选项

// Delay in seconds
await this.schedule(60, "taskMethod", { data });

// Specific date
await this.schedule(new Date("2025-01-01T00:00:00Z"), "taskMethod", { data });

// Cron expression (recurring)
await this.schedule("0 9 * * *", "dailyTask", {});  // 9 AM daily
await this.schedule("*/5 * * * *", "everyFiveMinutes", {});  // Every 5 min

// Manage schedules
const schedules = await this.getSchedules();
await this.cancelSchedule(taskId);

聊天 Agent（AI 驱动）

对于专注于聊天的 Agent，可扩展 AIChatAgent：

import { AIChatAgent } from "@cloudflare/ai-chat";

export class ChatBot extends AIChatAgent<Env> {
  // Called for each user message
  async onChatMessage(message: string) {
    const response = await this.env.AI.run("@cf/meta/llama-3-8b-instruct", {
      messages: [
        { role: "system", content: "You are a helpful assistant." },
        ...this.messages,  // Automatic history management
        { role: "user", content: message },
      ],
      stream: true,
    });

    // Stream response back to client
    return response;
  }
}

包含的功能：

• 自动保存消息历史
• 可恢复的流式传输（断开连接后仍可继续）
• 内置 saveMessages()，用于持久化

客户端集成

React Hook

import { useAgent } from "agents/react";

function Chat() {
  const { state, send, connected } = useAgent({
    agent: "my-agent",
    name: userId,  // Agent instance ID
  });

  const sendMessage = (text: string) => {
    send(JSON.stringify({ type: "chat", content: text }));
  };

  return (
    <div>
      {state.messages.map((msg, i) => (
        <div key={i}>{msg.role}: {msg.content}</div>
      ))}
      <input onKeyDown={(e) => e.key === "Enter" && sendMessage(e.target.value)} />
    </div>
  );
}

原生 JavaScript

const ws = new WebSocket("wss://my-agent.workers.dev/agents/MyAgent/user123");

ws.onopen = () => {
  console.log("Connected to agent");
};

ws.onmessage = (event) => {
  const data = JSON.parse(event.data);
  console.log("Received:", data);
};

ws.send(JSON.stringify({ type: "chat", content: "Hello!" }));

常见模式

请参阅 references/agent-patterns.md，了解：

• 工具调用和函数执行
• 多 Agent 编排
• RAG（检索增强生成）
• 人在回路工作流

部署

# Deploy
npx wrangler deploy

# View logs
wrangler tail

# Test endpoint
curl https://my-agent.workers.dev/agents/MyAgent/test-user

故障排除

常见问题请参阅 references/troubleshooting.md。

参考资料

• references/examples.md — 官方模板和生产环境示例
• references/agent-patterns.md — 高级模式
• references/state-patterns.md — 状态管理策略
• references/troubleshooting.md — 错误解决方案