Cloudflare Launches Cloudflare Computer for AI Agents

Cloudflare has introduced Cloudflare Computer, an open-source runtime environment designed to run persistent, stateful AI agents. Available via the @cloudflare/computer package, this new infrastructure aims to replace traditional, resource-heavy container setups with a hybrid model. By leveraging lightweight serverless isolates alongside container sandboxes, the system allows agents to run faster and scale more cost-effectively than they would on standard virtual machines.

The company argues that relying solely on containers to run agents cannot scale to what Cloudflare describes as "hundreds of millions, then billions, of concurrent agents" due to global compute limitations. To solve this, Cloudflare Computer dynamically decides whether to execute code in an isolate, a container sandbox, or a web browser. Cloudflare aims for a distribution where a container is required for less than 10% of an agent's workload, leaving tasks like document creation, coding, and audio or video manipulation to faster, cheaper isolates.

Under the hood, the architecture runs the main agent harness inside a Durable Object isolate and calls an attached container on-demand as a tool. A shared, SQLite-based filesystem connects the two environments, allowing tasks to transition smoothly between isolates and containers while working on the same files. This filesystem can integrate with git repositories, storage buckets, or arbitrary files, with all operations audited and gated.

Currently, the platform offers three backends: container projects that expose the SQLite state as a FUSE-mounted filesystem, isolate shells running a bash environment in a Dynamic Worker, and isolate JavaScript running an ECMAScript module. For developers, this means a shift toward highly scalable, hybrid agent hosting, though Cloudflare notes that the project is currently in an early preview phase suitable only for experiments and prototypes.