A quiet but significant announcement came out of Cloudflare this week. The company introduced @cloudflare/computer, an open-source package that gives AI agents their own virtual computer to work with, complete with a filesystem, shell access, and the ability to run code across multiple execution environments simultaneously. The tagline from Cloudflare’s blog puts it simply: “Your agent needs a computer, not a container.” It sounds like developer tooling jargon. But the idea behind it touches on one of the most pressing infrastructure problems in AI right now, and for founders building AI products in the Middle East and beyond, understanding it matters more than it might initially appear. The Problem That Nobody Is Talking About Loudly Enough To understand what Cloudflare built, you first need to understand why the current approach to running AI agents at scale is running into a wall. Most capable AI agents today work by being given access to a computer: a filesystem, a shell, the ability to install packages and run code. Coding agents like Claude Code, Cursor, and similar tools work this way. You give the model an environment to operate in, and it inspects that environment, makes changes, tests its work, and keeps going until the task is done. The standard way to provide that environment has been containers, isolated Linux environments that give each agent its own slice of compute. This works fine for small deployments. But Cloudflare’s engineers have identified a fundamental scaling problem: there is not enough compute on the planet to give every user’s agent its own containerized environment if AI agents go mainstream. Think about what that means in practice. If every enterprise deploys AI agents for every employee, and every consumer app gives users personal agents to handle tasks, the number of concurrent agent environments needed reaches hundreds of millions, then billions. Running a full container for each one is computationally prohibitive. It’s not just expensive. It’s physically impossible given current infrastructure. McKinsey reports that about 62% of organizations are already experimenting with or piloting AI agents, with 23% scaling them in at least one business function. AI workloads are projected to make up 71% of data center demand by 2030. The gap between current container-based approaches and the compute those projections imply is enormous. This is why Cloudflare framed the problem explicitly: “Across all the clouds, all the hyperscalers, there’s nowhere near enough compute in the world for every company to give each of their users’ agents their own containerized compute environment.” Cloudflare’s Bet: Isolates, Not Containers Cloudflare’s answer draws on a bet the company made almost ten years ago when it launched Cloudflare Workers. Workers use a different model than containers. Instead of spinning up a full operating system environment for each workload, they use “isolates”: lightweight JavaScript execution environments that start in milliseconds, use a fraction of the memory a container requires, and can scale horizontally to millions of concurrent instances. The trade-off has always been capability. Isolates are fast and cheap but can’t run arbitrary Linux programs or install native binaries. Containers are more capable but heavy, slow to start, and resource-intensive. @cloudflare/computer’s insight is that most of what an AI agent actually does doesn’t require a full container. Reading files, writing code, processing data, managing git repositories, manipulating text, these tasks can run in lightweight isolates. Only specific operations, installing npm packages, running native binaries, executing Linux commands, require a full container environment. So @cloudflare/computer gives an agent both. A shared filesystem sits at the center, backed by SQLite and accessible from both execution environments. The agent can run fast, cheap isolate operations for the majority of its work and spin up a container only when it genuinely needs one. Cloudflare’s own benchmarks aim for containers being required for less than 10% of an agent’s tasks. The result is an architecture where the agent harness (the reasoning loop) runs in a Durable Object, infinitely scalable horizontally, while the container provides vertical compute on demand only when necessary. Horizontal scale for the intelligence. Vertical scale only when the task demands it. What Developers Actually Get For developers building AI-powered products, @cloudflare/computer provides a few concrete things. A virtual filesystem that works like a real one: readable, writable, and editable by the agent with built-in git support so you can clone repositories directly into the workspace. All operations are gated, audited, and observable, giving developers a clear record of every action the agent took. An AI toolkit with standard tools: read, write, edit, li
