Cloudflare Unveils @cloudflare/computer for Scalable Agent Runtimes
Cloudflare introduces @cloudflare/computer, an open-source library abstracting compute primitives to provide agents with a unified 'computer' environment, aiming to overcome containerization scalability limits.
Cloudflare has announced an early preview of @cloudflare/computer, an open-source library designed to serve as an agent runtime. This new offering aims to provide agents with a unified 'computer' environment, abstracting away the complexities of underlying compute primitives such as isolates and containers. The core idea is to offer a more scalable and efficient solution for the burgeoning demand of agentic systems, moving beyond the limitations of traditional containerization.
The motivation behind @cloudflare/computer stems from the observation that capable agents often perform best when provided with their own dedicated computing environment, complete with a filesystem, tools, and the ability to execute code. However, scaling this model to potentially billions of concurrent agents presents a significant challenge. The current industry trend of relying on containerization for agent execution faces a fundamental scalability issue, as the world's available compute resources are insufficient to provide each agent with its own container.
Cloudflare's long-standing investment in isolates, pioneered with Cloudflare Workers nearly a decade ago, offers a more promising path. Isolates are designed for infinite horizontal scalability, rapid spin-up and tear-down times, and the ability to hibernate when idle. This makes them an ideal primitive for agentic workloads that require rapid scaling and efficient resource utilization. The company has further enhanced isolates by enabling them to spin up container sandboxes on demand, allowing for the use of heavier compute primitives only when necessary.
@cloudflare/computer builds upon this foundation by providing a unified abstraction layer. It allows developers to define a primed filesystem declaratively, containing all necessary components for a given task. Agents can then leverage different execution environments—either lightweight isolates for file manipulation and data processing, or more robust containers for tasks requiring Linux environments or native binaries—all operating on the same synchronized filesystem.
The library introduces a durable filesystem backed by SQLite, which can be populated from various sources like cloud storage or source control. It also includes tools for reading, writing, and editing files using code or bash commands, with all operations being gated, audited, and observed for fine-grained control and accountability. This ensures that agents can perform tasks securely and that a clear audit trail of their actions is maintained.
Users can instantiate a @cloudflare/computer workspace on any Durable Object, providing a virtual filesystem and execution runtime. The package is installable via npm. The primary use case involves equipping agents with this filesystem and tooling. For instance, an agent designed to triage bug reports could utilize a workspace instance. The package offers several execution backends, including Cloudflare Containers, and allows for custom backend development. Tools for file manipulation, git operations, and shell access can be exposed alongside product-specific tools.
While the model can use these tools during its agent loop, developers can also interact with the workspace API directly, for example, to pre-configure the environment before initiating the agent. The workspace repository provides further examples and a step-by-step tutorial for building agents from scratch, demonstrating the flexibility and power of this new abstraction.
By open-sourcing @cloudflare/computer, Cloudflare aims to collaborate with customers who are pushing the boundaries of running agents at scale. This initiative seeks to gather feedback and refine the library, ultimately contributing to more efficient and scalable agentic systems that can meet the demands of the rapidly evolving AI landscape.