Cloudflare has released Project Think to enable AI agents to be deployed to the cloud like websites: self-recovering from crashes, with zero cost during idle time.
According to Dynamic Beating monitoring, Cloudflare has released Project Think, adding a comprehensive set of infrastructure for building long-lived AI agents to its open-source Agents SDK: persistent execution, subagents, sandboxed code execution, and persistent sessions. Developers can use these components individually or quickly build a complete agent application using the Think base class and deploy it to the Cloudflare global network with a single npx wrangler deploy command. Currently in preview.
The core issue Project Think aims to address is that current programming agents (such as Claude Code, Codex, OpenClaw, etc.) can only run on local laptops or expensive VPS, are disrupted when a laptop is closed, cannot support multi-user collaboration, and continue to incur costs when idle. Cloudflare's solution is based on Durable Objects, where each agent is an independent actor with a built-in SQLite database. When idle, the agent sleeps and incurs zero cost. According to Cloudflare's calculations, if 10,000 agents are each active only 1% of the time, a traditional container solution would require 10,000 resident instances, while Durable Objects would have only about 100 running at any given time.
There are several key design aspects worth noting:
1. Persistent Execution (Fibers): If an agent's environment crashes during runtime (due to deployment updates, platform restarts, resource exhaustion), the SDK records a checkpoint in SQLite. Upon restart, the agent resumes from the checkpoint without losing progress.
2. Code Mode (codemode): Instead of having a model make individual API calls and read results one by one, Cloudflare's approach allows the model to write a complete program to accomplish the task at once. For example, using Cloudflare's MCP API server as an example: exposing all API endpoints would require about 1.17 million token descriptions, but with codemode, only two tools (search and execute) totaling about 1000 tokens are needed, reducing by 99.9%.
3. Five-Tier Execution Model: Ranging from the lightest virtual file system (Tier 0) to a full sandboxed environment (Tier 4) supporting operations like git clone and npm test, agents can incrementally upgrade their capabilities on-demand without having to start with heavyweight containers.
4. Self-Written Extensions: Agents can dynamically write TypeScript extensions at runtime and register them as new tools. These extensions run in a sandboxed environment with controlled permissions. For instance, if a user asks an agent to manage GitHub PRs, the agent can instantly write a GitHub integration extension and call it directly next time.
Cloudflare categorizes AI agents into three stages in their blog: the first wave is stateless chatbots, the second wave is locally run programming agents, and the third wave is persistent agents running as cloud infrastructure. Project Think represents Cloudflare's bet on the third wave. For developers, the appeal of this solution lies in shifting the operational cost of agents from "pay per instance" to "pay per actual usage." If an agent spends most of its time waiting for user commands, the cost approaches zero.
