Cloudflare’s @cloudflare/computer: An Agent Runtime That Ditches Containers for Isolates

Cloudflare has open-sourced @cloudflare/computer, an early-preview agent runtime built on its isolate-based Workers platform. The pitch: agents don’t need a container per instance — they need a computer, and the platform should decide where code actually runs.

The package provides a durable virtual filesystem (backed by SQLite) that agents can read, write, and execute against. Execution can happen inside a lightweight isolate or, when the task needs a full Linux userland (npm, native binaries), inside an attached container. Both operate on the same filesystem, kept in sync. The agent picks the right backend for the job, and the runtime handles the plumbing.

This is a direct challenge to the current norm of spinning up a container for every agent session. Cloudflare argues that approach won’t scale to the projected hundreds of millions or billions of concurrent agents — there simply isn’t enough container compute in the world. Isolates, which Cloudflare has bet on for nearly a decade, spin up and tear down in milliseconds and hibernate when idle, making them the only realistic path to horizontal scale for agent workloads.

The library is designed to plug into Cloudflare’s @cloudflare/think agent harness. You instantiate a Workspace on a Durable Object, wire up backends (like CloudflareContainerBackend), and expose file, git, and shell tools to the model. All operations are gated, audited, and observed, giving developers a paper trail of what the agent did.

Cloudflare is shipping this as an experiment, open source, to learn with customers who are pushing agent scale. The core idea — separate the “brain” (agent loop) from the “hands” (execution sandbox) — is already gaining traction, but Cloudflare’s twist is making the sandbox itself adaptive: isolate for file ops, container for heavy lifting, all against a shared filesystem.

For engineers building agents, this is a meaningful alternative to the container-per-agent model. It’s early, and the package is an early preview, but the direction is clear: the future of agent compute is heterogeneous, and the runtime should hide that complexity.