Cloudflare Computer
TLDR
Cloudflare launched an open-source agent runtime that gives AI agents a shared filesystem across lightweight isolates and full container sandboxes. It's a bet that most agent work doesn't need a full Linux box, which could make running millions of agents actually affordable.
Cloudflare just open-sourced something it's calling @cloudflare/computer, and the pitch is refreshingly narrow: agents need a computer, but they don't all need the same kind of computer. The package hands an agent a durable, SQLite-backed virtual filesystem plus a choice of places to actually run code against it — either a lightweight V8 isolate or a full Cloudflare Container with a real Linux userland. Both stay in sync with the same files, so an agent can edit a config in the cheap isolate and then drop into a container only when it needs npm, a native binary, or something that isolates simply can't do.
The reasoning behind this is more interesting than the plumbing. Cloudflare argues that giving every single agent its own containerized environment just doesn't scale — not for hundreds of millions of concurrent agents, and definitely not for the billions that are supposedly coming. That's the polite version of what the company calls a 'desperate, panicked industry demand for CPU compute.' Cloudflare has been quietly building toward this moment for almost a decade, first with Workers in 2017, then Durable Objects around 2019, both bets on isolates being cheaper, faster to spin up, and more horizontally scalable than containers ever will be.
What's new isn't the isolate-versus-container split itself — Cloudflare already let Durable Objects call containers on demand as tools. What's new is bundling that pattern into a single library so developers stop hand-rolling the glue themselves. Under the hood, the isolate runtime translates shell commands into JavaScript via something called just-bash, while the container runtime mounts the same files through FUSE so changes sync both ways. Developers wire it up with a handful of lines — install the npm package, attach a Workspace to a Durable Object, hand the agent read/write/edit/exec tools — and, notably, Cloudflare says frontier models are already pretty good at picking the right backend on their own without much prompting.
Cloudflare's stated goal is blunt: get containers down to handling less than 10% of an agent's actual workload, with isolates covering coding, document generation, and even audio or video manipulation. That's a big ask given how much of today's agent tooling assumes a full sandboxed OS by default. But if it holds up outside Cloudflare's own dogfooding — bug-triage agents, deployment bots, and browser-driving tasks are the early examples cited — it reframes the container-per-agent model as the expensive exception rather than the default.
My take
This is Cloudflare doing what it always does: turning a scaling problem everyone else is throwing hardware at into an architecture argument, and conveniently one where its decade-old isolate bet finally pays off. The 90/10 split they're chasing is the right instinct — most agent work really is just file-shuffling, not compiling a kernel — but expect every other cloud vendor to ship a near-identical 'lightweight-by-default' story within a year, because nobody wants to be the one still selling containers by the gigabyte-second when the market decides that's the wasteful option.
Read more about this at: TLDR
Related stories
AWS, Google Cloud, Microsoft Azure, and Cloudflare now all offer agent sandboxes. None built them the same way.
The New Stack · 1 week ago ·
46
The rise of the agent runtime: The compute platform behind production agents
The New Stack · 2 weeks ago ·
19
Enterprises power agentic workflows in Cloudflare Agent Cloud with OpenAI
OpenAI Blog · 3 months ago ·
47