Cloudflare ships @cloudflare/computer, an isolates-first agent runtime

Cloudflare's pitch for its new agent runtime, laid out in a post on the Cloudflare blog by Matt Carey and Aron Carroll, is that the industry's default assumption is already wrong. The assumption is that every agent gets its own container, the way every user session used to get its own VM. Their counter-claim is blunt: "there's nowhere near enough compute in the world for every company to give each of their users' agents their own containerized compute environment."

The answer they are shipping is an open source npm package, `@cloudflare/computer`, that treats isolates as the base unit and containers as a specialized tool the agent reaches for only when it has to. Isolates, the same primitive that has powered Cloudflare Workers for nearly a decade, are described as "infinitely horizontally scalable" and quick to spin up and tear down. The stated target is a runtime where "a container is required for less than 10% of its work."

The piece that makes the model actually usable is what they call the workspace: a virtual filesystem backed by SQLite that can be populated from cloud storage or source control, and that stays in sync when an agent has to escalate into a full Linux container via a FUSE mount. Inside the isolate side, shell commands are handled by a tool called just-bash, which translates shell code into JavaScript so it can run in a worker. Developers get an AI SDK compatible toolkit exposing the usual moves: read, write, edit, ls and exec.

The honest caveat is that this is an early preview and the post is a first-party pitch. Translating bash to JavaScript is the sort of abstraction that works beautifully in demos and then leaks in production on the one flag your agent actually needed, and the excerpt does not characterise how the SQLite workspace behaves under heavy concurrent writes or how long a container escalation actually takes. Pricing, quotas, and whether any of this runs outside Cloudflare's own platform are not addressed.

The interesting part is what it means for everyone else. If Cloudflare is right that the container-per-agent model breaks at scale, then the current crop of agent-sandbox vendors has a real architectural argument to answer, and teams building coding, research, and browser-automation agents get a plausible cheaper on-ramp to try. Worth reading as a bet on where the cost curve goes, not as a settled result.