What Cloudflare Announced
Cloudflare has introduced Cloudflare Computer, an open-source runtime designed to give AI agents something closer to a real computer rather than a short-lived container. The company says the system uses Cloudflare isolates for fast serverless execution, with the aim of making agents cheaper to run, faster to start, and easier to scale.
In this context, an AI agent is software that can keep context, call tools, and complete tasks across multiple steps. A runtime is the execution layer that provides compute, state, files, and resource management. Cloudflare Computer’s central idea is that each agent should have a persistent environment that can sleep when idle and resume when needed.
Why Containers Are Not the Default Answer
Cloudflare frames the launch around a scaling problem. Running agents in containers may work for many current applications, but the company argues that this approach cannot scale to “hundreds of millions or even billions” of concurrent agents because there is not enough global compute capacity to support that model efficiently.
Containers are useful because they provide strong isolation and a familiar system environment. But they also carry overhead in startup time, scheduling, and resource consumption. For many agent tasks—such as maintaining state, executing lightweight code, or coordinating tool calls—a full container may be more than is necessary.
The @cloudflare/computer package introduces a runtime where the platform decides whether code should run in an isolate, a container sandbox, or a web browser. An isolate is a lightweight isolated execution environment commonly used in serverless and edge computing. Cloudflare says the isolates introduced with Cloudflare Workers can scale horizontally and start or stop very quickly.
Cloudflare’s target is that less than 10% of agent work should require containers, while tasks such as coding, audio and video processing, and document creation can be handled by isolates. In this model, containers become heavier tools that are invoked only when needed.
The Architecture: Isolates, Containers, and Shared Files
Cloudflare Computer runs the agent framework inside an isolate, specifically through a Durable Object. A Durable Object is Cloudflare’s mechanism for stateful compute, allowing a specific object to preserve state and coordinate activity in a distributed environment. An agent can keep its state in the isolate, sleep when it is not active, and start its own container sandbox when a task requires a heavier execution environment.
This design combines horizontal and vertical scaling. Isolates handle lightweight, high-concurrency work, while containers provide a more complete environment for heavier operations. Cloudflare describes the containers as on-demand tools, so developers pay the cost of a heavier compute primitive only when it is necessary.
A key component is a shared file system based on SQLite. SQLite is an embedded relational database often used for local storage. In Cloudflare Computer, it serves as the basis for shared state and files that both isolates and containers can access. This lets a task move between the two execution environments without rebuilding context or copying files into a separate workspace.
According to Cloudflare, the file system can work with:
- Git repositories;
- storage buckets;
- arbitrary files.
The company also emphasizes that operations are governed, audited, and observable. Observability matters for agent systems because developers need to understand what actions were taken, what tools were invoked, and how state changed during multi-step workflows.
Current Backends and Preview Status
Cloudflare Computer currently offers three backends. The first is a container project that exposes SQLite state to sandboxed containers as a real FUSE-mounted file system. FUSE is a userspace file system mechanism that allows applications to implement file system behavior without modifying the operating system kernel. The second backend is an isolate shell, which runs a just-bash environment inside a Dynamic Worker. The third is isolate JavaScript, which runs ECMAScript modules inside a new Dynamic Worker.
Together, these backends show that Cloudflare Computer is not simply another sandbox. It is a runtime and state abstraction that can shift work across lightweight JavaScript execution, shell-like environments, and containers while preserving access to shared files.
Cloudflare is also clear about maturity: Cloudflare Computer is in early preview and is intended for experimentation, exploration, and prototyping. Developers should therefore treat it as an emerging platform rather than a stable production foundation.
Why It Matters
The launch highlights a broader shift in AI infrastructure. As agents move from chat interfaces to task execution, they need persistent state, file access, tool invocation, audit trails, and elastic compute. Sending every task into a container may be too expensive or too slow for large-scale agent deployments.
Cloudflare’s approach makes the lightweight isolate the default execution layer and uses containers only as needed. That matches the shape of many agent workloads: most steps may involve reading state, updating files, or running small pieces of logic, while only a minority need a full system environment. If this model proves practical, agent platforms may evolve from deploying applications to assigning each agent a persistent, observable, and suspendable computer. For now, Cloudflare Computer remains an early preview, but it points toward a more layered runtime model for future AI agents.