Building AI Agent On Cloudflare

Building Cloudflare Agents

Your knowledge of the Agents SDK may be outdated. Prefer retrieval over pre-training for any agent-building task. Always fetch from the sources below before writing agent code.

Retrieval Sources

When to Use

• User wants to build an AI agent or chatbot on Cloudflare.
• User needs stateful, real-time AI interactions.
• User asks about the Cloudflare Agents SDK.
• User wants scheduled tasks or background AI work.
• User needs WebSocket-based AI communication.
• User mentions "Agents SDK", "Durable Objects agent", or "edge AI agent".

Prerequisites

• Cloudflare account with Workers enabled.
• Node.js 18+ and npm/pnpm/yarn installed.
• Wrangler CLI installed globally: npm install -g wrangler.
• Windows host is primary — use PowerShell for all local commands. Paths use Windows conventions (~).

Procedure

1. Scaffold a New Agent Project

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

Agent runs locally at http://localhost:8787.

2. Define the Agent Class

Create src/agent.ts with a class extending Agent<Env, State>:

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) {
    const messages = [
      ...this.state.messages,
      { role: "user", content: userMessage },
    ];

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

    this.setState({
      ...this.state,
      messages: [
        ...messages,
        { role: "assistant", content: response.response },
      ],
    });

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

3. Configure the Entry Point

Create src/index.ts:

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

export default {
  async fetch(request: Request, env: Env) {
    return (
      (await routeAgentRequest(request, env)) ||
      new Response("Not found", { status: 404 })
    );
  },
};

export { MyAgent };

Clients connect via: wss://my-agent.workers.dev/agents/MyAgent/session-id

4. Configure Wrangler

Create or edit wrangler.toml:

name = "my-agent"
main = "src/index.ts"
compatibility_date = "2024-12-01"

[ai]
binding = "AI"

[durable_objects]
bindings = [{ name = "AGENT", class_name = "MyAgent" }]

[[migrations]]
tag = "v1"
new_classes = ["MyAgent"]

5. Manage State

Reading state — current state is always available via this.state:

const currentMessages = this.state.messages;
const userPrefs = this.state.preferences;

Updating state — 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 Storage — for complex queries, use the embedded SQLite database:

// 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}%`}
`;

6. Add Scheduled Tasks

Agents can schedule future work using this.schedule():

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

  if (data.type === "schedule_reminder") {
    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 }) {
  console.log(`Reminder for ${data.userId}: ${data.message}`);

  this.setState({
    ...this.state,
    lastReminder: new Date().toISOString(),
  });
}

Schedule options:

// 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);

7. Build a Chat Agent with AIChatAgent

For chat-focused agents, extend AIChatAgent instead of Agent:

import { AIChatAgent } from "agents/ai-chat-agent";

export class ChatBot extends AIChatAgent<Env> {
  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,
    });

    return response;
  }
}

Features included with AIChatAgent:

• Automatic message history.
• Resumable streaming (survives disconnects).
• Built-in saveMessages() for persistence.

8. Integrate Client-Side

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 => e.key === "Enter" && sendMessage(e.target.value)} />
    </div>
  );
}

Vanilla 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!" }));

9. Deploy

# Deploy to Cloudflare Workers
npx wrangler deploy

# View live logs
wrangler tail

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

Advanced Patterns

When to load reference files:

• references/agent-patterns.md — Load when the user asks for tool calling, function execution, multi-agent orchestration, RAG (Retrieval Augmented Generation), or human-in-the-loop workflows.
• references/state-patterns.md — Load when the user needs advanced state management strategies beyond basic setState or SQL.
• references/examples.md — Load when the user wants official templates or production-ready examples to start from.
• references/troubleshooting.md — Load when the user encounters errors during development, deployment, or runtime.

Pitfalls

• Outdated SDK knowledge: The Agents SDK evolves rapidly. Always retrieve from the GitHub docs or developers.cloudflare.com/agents/ before writing code. Do not rely on pre-trained knowledge alone.
• Missing Durable Object migration: If you add a new Agent class, you must add a [[migrations]] entry with new_classes. Without it, the Worker will fail to deploy or route correctly.
• State not syncing: setState must be called with a new object (not a mutation of this.state). Directly mutating this.state will NOT persist or sync to clients.
• SQL injection risk: Always use tagged template literals (this.sql\...``) with parameter placeholders. Never interpolate raw user input into SQL strings.
• Schedule method name mismatch: The method name passed to this.schedule() (e.g., "sendReminder") must exactly match a method on the Agent class. Typos cause silent failures.
• Model availability: @cf/meta/llama-3-8b-instruct is used in examples. Verify model availability in the user's Cloudflare account region before deploying. Check Workers AI model catalog for current models.
• Windows path issues: On Windows PowerShell, use backslash paths for local files. Wrangler and Node commands work the same, but file paths in scripts should use Windows conventions.
• routeAgentRequest not exported: Ensure MyAgent is re-exported from src/index.ts (export { MyAgent }). Without this export, the Durable Object binding will not resolve.
• Port conflicts: Local dev server defaults to http://localhost:8787. If that port is in use, npm start will fail. Check with netstat -ano | findstr 8787 in PowerShell.

Verification

1. Verify local server is running:

   curl http://localhost:8787
   

   Expected: a response (not a connection error).

2. Verify WebSocket connection:

   # Use a WebSocket client or browser console
   # Connect to: ws://localhost:8787/agents/MyAgent/test-session
   # Send: {"type":"chat","content":"Hello"}
   # Expect: {"type":"response","content":"..."}
   

3. Verify deployment:

   npx wrangler deploy
   curl https://my-agent.workers.dev/agents/MyAgent/test-user
   

   Expected: successful deploy output and a non-404 response.

4. Verify state persistence:

   • Connect to an agent session, send messages, disconnect.
   • Reconnect to the same session ID.
   • Confirm onConnect sends back the full message history.

5. Verify scheduled tasks:

   // In agent code, log when scheduled method fires
   console.log("Scheduled task fired");
   

   • Trigger a short-delay schedule (e.g., this.schedule(5, "testTask", {})).
   • Run wrangler tail and confirm the log appears after 5 seconds.

References

• references/examples.md — Official templates and production examples
• references/agent-patterns.md — Advanced patterns: tool calling, multi-agent, RAG, human-in-the-loop
• references/state-patterns.md — State management strategies
• references/troubleshooting.md — Error solutions and common issues