Cloudflare Computer Launched: Giving AI Agents a Persistent Runtime Environment

Cloudflare has introduced Cloudflare Computer, a new open-source runtime aimed at providing AI agents with a persistent runtime environment, moving beyond temporary containers. According to the company, it uses Cloudflare isolates to enable fast serverless execution, reducing costs and improving speed and scalability for agents. Cloudflare argues that container-based agent deployment cannot scale to 'hundreds of millions or even billions of concurrent agents' due to global compute limitations. The @cloudflare/computer package introduces an agent runtime where the platform decides whether code runs in an isolate, a container sandbox, or a web browser, giving each agent a computer while optimizing for efficiency and scalability. Cloudflare states that less than 10% of work requires containers, with isolates handling coding tasks, audio/video processing, and document creation. Isolates, launched alongside Cloudflare Workers, scale horizontally without limits and start and shut down quickly; they persist agent state, hibernate when idle, and can spawn container sandboxes when needed. The architecture runs the agent framework in an isolate (a Durable Object) and uses containers as on-demand tool calls, optimizing performance and cost. A core component is a SQLite-based shared file system accessible by both isolates and containers, allowing seamless task transfer and file handling. Cloudflare Computer currently offers three backends: container projects (exposing SQLite state via FUSE), isolate shell (running a just-bash environment in Dynamic Worker), and isolate JavaScript (running ECMAScript modules in a new Dynamic Worker). Cloudflare Computer is in early preview, suitable for experimentation, exploration, and prototyping.