Cloudflare Unveils AI Agent Runtime with Stateful Compute Environment

Cloudflare has unveiled Cloudflare Computer, an open-source runtime that moves AI agents beyond ephemeral container deployments toward persistent, stateful environments more akin to traditional computing. The platform harnesses Cloudflare's isolate architecture to achieve better economics and throughput.

The Scaling Challenge#

Existing container-based approaches face a fundamental constraint: deploying "hundreds of millions, then billions, of concurrent agents" requires more global compute infrastructure than currently exists. Cloudflare's thesis is that relying exclusively on containers becomes untenable at planetary scale, prompting a rethink of how agent runtimes should be architected.

Hybrid Execution Model#

Rather than forcing all workloads into a single execution tier, the @cloudflare/computer package lets the runtime intelligently distribute tasks across isolates, container sandboxes, and browser environments. Each agent receives its own logical "computer" where the platform optimizes placement: lightweight operations run in isolates, heavier jobs spawn containers on demand. Cloudflare targets a model where containers handle less than 10% of an agent's work, with coding tasks, media processing, and document generation handled by faster isolate execution.

Shared State and Filesystem#

Central to this architecture is a SQLite-backed filesystem layer that both isolates and containers can access. This allows work to flow seamlessly between execution tiers while operating on the same files—enabling git repositories, cloud storage, and arbitrary file operations while maintaining audit trails and access controls.

Available Backends#

The platform currently offers three execution options: container projects (exposing SQLite state as a FUSE-mounted filesystem), isolate shells (running a bash environment in a Dynamic Worker), and isolate JavaScript (executing ECMAScript modules in a fresh Dynamic Worker).

Cloudflare notes the technology remains in early preview, suited for experimentation and prototypes rather than production use at this stage.