Cloudflare has released Cloudflare Computer, an open-source runtime designed to give AI agents a persistent, stateful "computer" instead of ephemeral containers. The company argues that the container-based approach many agents rely on today will not scale to the "hundreds of millions, then billions, of concurrent agents" it expects in the near future. The announcement, published on Aug 07, 2026, positions the runtime as a way to make agents cheaper, faster, and more scalable by leaning on Cloudflare's isolate technology.
Why Containers Fall Short
Cloudflare's core argument is simple: containers are too heavy for most agent work. Each container spins up an entire operating environment, which consumes global compute capacity at a rate the company says is unsustainable. Instead, Cloudflare Computer uses isolates, the lightweight serverless execution model introduced with Cloudflare Workers. Isolates start and shut down extremely quickly, persist agent state, and can hibernate when the agent is not running. They are, in Cloudflare's words, "infinitely horizontally scalable."
The company's goal is that a container will be required for less than 10% of an agent's work. Coding tasks, audio and video manipulation, and document creation can all be handled by isolates, according to Cloudflare. That leaves containers for the rare, heavy compute jobs that genuinely need a full sandbox.
The Architecture: Isolates First, Containers On Demand
The runtime's design centers on a simple model: "each agent gets a computer, the runtime optimizes for efficiency, and scalability." The agent harness runs inside an isolate, hosted in a Durable Object, Cloudflare's storage and coordination primitive. When the agent needs heavier compute, that isolate can spin up its own container sandbox on demand, calling it as a tool.
This hybrid approach means the runtime only pays for container-level resources when absolutely necessary. Cloudflare says this optimizes both performance and cost. The @cloudflare/computer npm package implements the runtime, and it lets the platform decide whether any given piece of code runs in an isolate, a container sandbox, or a web browser. That flexibility is central to the design.
A Shared Filesystem Bridges the Gap
A key technical detail is the shared filesystem. Built on SQLite, it is accessible to both isolates and containers. Tasks can move seamlessly between the two environments using this common storage layer. The filesystem can be attached to git repositories, storage buckets, or arbitrary files, giving agents a consistent workspace regardless of where their code executes.
Stay ahead of the AI curve
The most important updates, news, and content — delivered weekly.
No spam. Unsubscribe anytime.
Cloudflare says all operations on this filesystem are gated, audited, and observed. That means every read and write is tracked, which the company frames as a safety feature for agent deployments. The shared filesystem is what allows an agent to start a task in an isolate, hand off a heavy computation to a container, and then pick up the results back in the isolate without losing state.
Three Backends and an Early Preview
The runtime currently ships with three backends. The first is container projects, where SQLite state is exposed as a sandboxed container with a real FUSE-mounted filesystem. The second is isolate shells, which run a just-bash environment inside a Dynamic Worker. The third is isolate JavaScript, which runs an ECMAScript module in a fresh Dynamic Worker.
Each backend serves a different purpose. Container projects are for tasks that genuinely need a full operating environment. Isolate shells are for quick command-line work. Isolate JavaScript is for lightweight script execution. Cloudflare says the combination of isolates and container sandboxes is extremely effective, though it did not provide independent benchmarks or external validation in the announcement.
Cloudflare Computer is still an early preview. The company is explicit that it is only suitable for experiments, exploration, and prototypes. That is a notable hedge, given the scale claims. The runtime is open source, so developers can inspect the code and test it, but Cloudflare is not yet recommending it for production agent workloads.
What This Means for Agent Development
Cloudflare's pitch is that agent infrastructure has been stuck in a container mindset, and that mindset will hit a wall as agent adoption grows. By moving most work into isolates, the runtime aims to cut costs and latency while enabling massive concurrency. The shared SQLite filesystem and on-demand container spawning are the mechanisms that make this possible.
The company says the architecture optimizes performance and cost, and that the shared filesystem allows tasks to move seamlessly. Whether those claims hold up in real-world, large-scale deployments remains to be seen, especially given the early preview status. But the direction is clear: Cloudflare is betting that the future of AI agents is not a container for every task, but a lightweight, persistent environment that reaches for a container only when it truly needs one.