- Published on
Cloudflare Launches Persistent, Stateful, Computer-like Environments for Agents
- Authors
- Name
- DeepHorus
Cloudflare has introduced Cloudflare Computer, an open-source runtime designed to give AI agents persistent, stateful environments closer to a real computer rather than short-lived containers. The platform leverages Cloudflare's isolates for fast serverless execution, aiming to make agents cheaper, faster, and more scalable. This launch directly addresses a growing bottleneck in AI agent deployment.
The company argues that relying primarily on containers to run agents will not scale to "hundreds of millions, then billions, of concurrent agents," because global compute capacity is simply insufficient. To overcome this, Cloudflare developed the @cloudflare/computer package, an agent runtime that decides whether code runs in an isolate, a container sandbox, or a web browser. In this model, each agent gets its own computer, while the runtime optimizes for efficiency and scalability.
According to Cloudflare, containers will be required for less than 10% of an agent's work, with coding tasks, audio/video manipulation, and document creation handled by isolates. Isolates, introduced with Cloudflare Workers, are described as "infinitely horizontally scalable" and can start up and shut down extremely quickly. They also persist agent state, hibernate when idle, and spin up their own container sandboxes when needed.
The architecture places the agent harness in an isolate within a Durable Object, while an attached container can be called on-demand as a tool. This hybrid approach combines horizontal scalability from isolates with vertical scalability from containers, optimizing both performance and cost. A shared, SQLite-based filesystem is accessible to both isolates and containers, enabling tasks to move seamlessly between them and work on the same files. Filesystems can integrate with git repositories, storage buckets, or arbitrary files, with all operations gated, audited, and observed.
Cloudflare Computer currently provides three backends: container projects, where SQLite state is exposed as a sandboxed, FUSE-mounted filesystem; isolate shells, which run a just-bash environment in a Dynamic Worker; and isolate JavaScript, which runs an ECMAScript module in a fresh Dynamic Worker. This flexibility allows developers to choose the right level of compute for each task.
The launch matters because it offers a practical path to massive concurrency for AI agents. Previous models would have consumed enormous container resources; Cloudflare's approach uses lightweight isolates for most work, reserving heavyweight containers for specific tasks. The ability to hibernate and resume agents also reduces idle time costs.
Cloudflare Computer remains an early preview, intended only for experiments, exploration, and prototypes. The company expects to refine the runtime based on developer feedback. If successful, this could reshape how AI agents are deployed, making large-scale autonomous workflows far more economical.
