Cloudflare Launches Project Think to Deploy AI Agents on Cloud-like Websites

ME News reports that on April 16 (UTC+8), according to monitoring by Beating, Cloudflare has launched Project Think, introducing a comprehensive infrastructure for building long-running AI agents to its open-source Agents SDK: persistent execution, sub-agents, sandboxed code execution, and persistent sessions. Developers can use these components individually or leverage the Think base class to rapidly build complete agent applications, deploying them instantly to Cloudflare’s global network with a single command: npx wrangler deploy. The offering is currently in preview. Project Think addresses a core challenge: current programming agents (such as Claude Code, Codex, OpenClaw, etc.) can only run on local laptops or expensive VPSs—laptops shut down when closed, collaboration among multiple users is impossible, and costs continue to accrue even during idle periods. Cloudflare’s solution is built on Durable Objects, where each agent is an independent actor equipped with its own SQLite database. It sleeps when idle and wakes upon receiving requests, resulting in zero computational cost during inactivity. According to Cloudflare’s calculations, if 10,000 agents are each active only 1% of the time, traditional container solutions would require 10,000 always-on instances; with Durable Objects, only about 100 are running at any given moment. Key technical design innovations include: 1. Persistent Execution (Fibers): If the environment crashes during agent operation—due to deployment updates, platform restarts, or resource exhaustion—the SDK records checkpoints in SQLite. Upon restart, the agent resumes from the last checkpoint without losing progress. 2. Code Mode (codemode): Instead of having the model call tools one by one and read results sequentially, codemode enables the model to write a complete program in one go to accomplish the task. Using Cloudflare’s own MCP server API as an example: exposing all API endpoints requires approximately 1.17 million tokens to describe tools; with codemode, only two tools (search and execute) are needed—roughly 1,000 tokens total—a 99.9% reduction. 3. Five-Tier Execution Hierarchy: Agents can incrementally upgrade capabilities from the lightest tier (Tier 0: virtual filesystem) to a full sandbox environment (Tier 4: supports git clone, npm test, etc.), scaling up only as needed rather than launching heavyweight containers from the start. 4. Self-Written Extensions: Agents can dynamically write and register TypeScript extensions at runtime as new tools. These extensions run within a sandbox with controlled permissions. For example, if a user instructs an agent to manage GitHub pull requests, the agent can immediately generate a GitHub integration extension and reuse it in future interactions. Cloudflare’s blog categorizes AI agents into three phases: the first wave consists of stateless chatbots; the second wave comprises locally running programming agents; the third wave involves persistent agents operating as cloud infrastructure. Project Think represents Cloudflare’s bet on this third wave. For developers, this solution is compelling because it shifts agent operational costs from “per-instance billing” to “pay-for-actual-use.” If an agent spends most of its time waiting for user commands, its cost approaches zero. (Source: BlockBeats)