Cloudflare Project Think: Build Long-Running AI Agents

Project Think is Cloudflare's new opinionated base class for building durable, long-running AI agents. Released during Agents Week 2026, it bundles persistent workspaces, sandboxed code execution, sub-agents, and durable sessions behind one Think class.

In this 5-minute guide, you'll spin up a Think agent with three commands, wire it to a model, and chat with it through a built-in WebSocket UI.

---

What you'll build

• A live Cloudflare Worker running a Think agent
• A persistent workspace where the agent reads and writes files
• A WebSocket chat UI talking to it in real time

---

Prerequisites

• Node.js 20+ installed
• A free Cloudflare account
• Wrangler CLI (installed by the starter template)

---

Step 1: Scaffold the starter

Run the official starter. It scaffolds a Worker, the Think class, and a React chat UI.

npm create cloudflare@latest -- \
  --template cloudflare/agents-starter \
  my-think-agent

cd my-think-agent
npm install

Expected output:

✔ Created my-think-agent
✔ Installed dependencies
→ Run `npm start` to launch the dev server

---

Step 2: Define your agent

Open src/agent.ts. The minimal Think agent only needs a getModel() method — Think handles the chat loop, persistence, and tools.

import { Think } from "@cloudflare/agents/think";
import { createWorkersAI } from "workers-ai-provider";

export class MyAgent extends Think {
  async getModel() {
    const ai = createWorkersAI({ binding: this.env.AI });
    return ai("@cf/meta/llama-3.3-70b-instruct");
  }
}

That's the entire agent. Think wires up: WebSocket chat protocol, message persistence, the agent loop, stream resumption, and workspace file tools.

---

Step 3: Run it locally

npm start

Open http://localhost:5173. Try a prompt:

You: Create notes/ideas.md with three startup ideas.

Agent: Created notes/ideas.md
  1. Voice-first inbox triage
  2. Auto-generated runbooks from incidents
  3. AI-driven feature flag rollouts
File saved to workspace ✓

Refresh the page — the file is still there. That's the persistent workspace.

---

Step 4: Understand the execution ladder

Think gives your agent five tiers of compute. Pick the lowest tier that does the job — they get more powerful but heavier as you climb.

Tier	Environment	Use it for
0	Workspace filesystem	Reading and writing files
1	Sandboxed JavaScript	Quick math, JSON parsing, transforms
2	JS + runtime npm	Running a library on the fly
3	Headless browser	Scraping, testing, screenshots
4	Full Linux sandbox	Shell, Python, long-running jobs

---

Step 5: Add a sub-agent

Sub-agents let the main agent delegate a subtask without bloating its own context.

import { Think, subAgent } from "@cloudflare/agents/think";

export class MyAgent extends Think {
  tools = {
    research: subAgent({
      name: "researcher",
      instructions: "Search the web and return a 5-bullet summary."
    })
  };

  async getModel() {
    return this.env.AI.run("@cf/meta/llama-3.3-70b-instruct");
  }
}

Now your agent can call research("...") and get back a clean summary instead of dumping raw search results into its main loop.

---

Step 6: Deploy to Cloudflare

npx wrangler deploy

Expected output:

✓ Built worker
✓ Uploaded to my-think-agent
→ https://my-think-agent.<your-subdomain>.workers.dev

Your agent now runs on Cloudflare's global network with state synced across requests.

---

Common pitfalls

• Forgetting the AI binding in wrangler.toml — add [ai] binding = "AI".
• Using Tier 4 sandbox for trivial work — expensive. Start at Tier 1.
• Storing secrets in workspace files — use wrangler secret put instead.

---

What's next

1. Add custom tools — define them in the tools object on your Think class.
2. Wire in a frontier model via the AI Gateway instead of Workers AI.
3. Explore the 30+ examples in the cloudflare/agents repo.

Project Think is in preview today. The full docs live at developers.cloudflare.com/agents.