agents-sdk

Cloudflare Agents SDK

Build persistent, stateful AI agents on Cloudflare Workers using the agents npm package. Prefer current Cloudflare Agents SDK docs over memory for API details because the SDK changes quickly.

FIRST: Verify Project And Installation

Inspect before mutating dependencies:

ls package.json wrangler.jsonc wrangler.toml 2>/dev/null
node -e "const p=require('./package.json'); console.log(p.dependencies?.agents || p.devDependencies?.agents || 'agents not installed')" 2>/dev/null

Install only when creating or updating an Agents project:

npm install agents

Agents require a binding in wrangler.jsonc:

{
  "durable_objects": {
    // "class_name" must match your Agent class name exactly
    "bindings": [{ "name": "Counter", "class_name": "Counter" }]
  },
  "migrations": [
    // Required: list all Agent classes for SQLite storage
    { "tag": "v1", "new_sqlite_classes": ["Counter"] }
  ]
}

Retrieval Rules

Check current Cloudflare Agents SDK documentation before using uncertain APIs, decorators, package names, Code Mode details, or wrangler migration syntax. Read existing wrangler.jsonc/wrangler.toml, lockfile, framework files, and compatibility date before editing.

Choosing an Agent Type

Key Concepts

• Agent base class provides state, scheduling, RPC, MCP, and email capabilities
• AIChatAgent adds streaming chat with automatic message persistence and resumable streams
• Code Mode generates executable code instead of tool calls—reduces token usage significantly
• this.state / this.setState() - automatic persistence to SQLite, broadcasts to clients
• this.schedule() - schedule tasks at Date, delay (seconds), or cron expression
• @callable decorator - expose methods to clients via WebSocket RPC

Quick Reference

Minimal Agent

import { Agent, routeAgentRequest, callable } from "agents";

type State = { count: number };

export class Counter extends Agent<Env, State> {
  initialState = { count: 0 };

  @callable()
  increment() {
    this.setState({ count: this.state.count + 1 });
    return this.state.count;
  }
}

export default {
  fetch: (req, env) => routeAgentRequest(req, env) ?? new Response("Not found", { status: 404 })
};

Streaming Chat Agent

Use AIChatAgent for chat with automatic message persistence and resumable streaming.

Install additional dependencies first:

npm install @cloudflare/ai-chat ai @ai-sdk/openai

Add wrangler.jsonc config (same pattern as base Agent):

{
  "durable_objects": {
    "bindings": [{ "name": "Chat", "class_name": "Chat" }]
  },
  "migrations": [{ "tag": "v1", "new_sqlite_classes": ["Chat"] }]
}

import { AIChatAgent } from "@cloudflare/ai-chat";
import { routeAgentRequest } from "agents";
import { streamText, convertToModelMessages } from "ai";
import { openai } from "@ai-sdk/openai";

export class Chat extends AIChatAgent<Env> {
  async onChatMessage(onFinish) {
    const result = streamText({
      model: openai("gpt-4o"),
      messages: await convertToModelMessages(this.messages),
      onFinish
    });
    return result.toUIMessageStreamResponse();
  }
}

export default {
  fetch: (req, env) => routeAgentRequest(req, env) ?? new Response("Not found", { status: 404 })
};

Client (React):

import { useAgent } from "agents/react";
import { useAgentChat } from "@cloudflare/ai-chat/react";

const agent = useAgent({ agent: "Chat", name: "my-chat" });
const { messages, input, handleSubmit } = useAgentChat({ agent });

Detailed References

• references/state-scheduling.md - State persistence, scheduling, queues
• references/streaming-chat.md - AIChatAgent, resumable streams, UI patterns
• references/codemode.md - Generate code instead of tool calls (token savings)
• references/mcp.md - MCP server integration
• references/email.md - Email routing and handling

When to Use Code Mode

Code Mode generates executable JavaScript instead of making individual tool calls. Use it when:

• Chaining multiple tool calls in sequence
• Complex conditional logic across tools
• MCP server orchestration (multiple servers)
• Token budget is constrained

See references/codemode.md for setup and examples.

Best Practices

1. Prefer streaming: Use streamText and toUIMessageStreamResponse() for chat
2. Use AIChatAgent for chat: Handles message persistence and resumable streams automatically
3. Type your state: Agent<Env, State> ensures type safety for this.state
4. Use @callable for RPC: Cleaner than manual WebSocket message handling
5. Code Mode for complex workflows: Reduces round-trips and token usage
6. Schedule vs Queue: Use schedule() for time-based, queue() for sequential processing
7. Validate config: Run wrangler types and project tests after binding or migration changes

Routing

• Use durable-objects for raw Durable Object design, storage, alarms, WebSockets, and RPC methods without the Agents SDK.
• Use this skill when the code uses Agent, AIChatAgent, McpAgent, agents, @cloudflare/ai-chat, persistent agent state, or Code Mode.