Say hello to agentOS (beta) A portable open-source OS built just for agents. Powered by WASM & V8 isolates. 🔗 Embedded in your backend ⚡ ~6ms coldstarts, 32x cheaper than sbxs 📁 Mount anything as a file system (S3, SQLite, …) 🥧 Use Pi, Claude Code/Codex/Amp/OpenCode soon

Why WebAssembly + V8? Traditional sandboxes boot entire VMs or containers. agentOS runs agents inside lightweight VMs within your own process. Same isolation technology behind Google Chrome. Battle-tested at planet scale. No VMs. No containers. No Docker images to pull. Just fast, secure execution.

The numbers: Cold start (p99): agentOS: 6.1 ms Fastest sandbox: 3,150 ms → 516x faster Memory per coding agent: agentOS: ~131 MB Sandbox minimum: ~1,024 MB → 8x less RAM Cost per execution-second: agentOS (AWS on-demand): $0.00000058/s Cheapest mainstream sandbox: $0.000018/s → 32x cheaper You're not paying for idle RAM anymore. Full methodology:

agentOS embeds directly into your existing backend. No separate infra. No sidecar. No Kubernetes operator. $ npm install rivetkit Your APIs. Your auth. Your toolchains. Agents run inside your own process, where your code already lives. A few lines of TypeScript and you're running agents.

Agents think in files. agentOS mounts anything as a familiar directory tree: - S3 buckets - SQLite databases - Google Drive - Host file system The host handles credential scoping. Agents never see API keys or secrets. They just read and write files like they always do.

Security model: - Granular CPU & memory limits per agent - Programmatic network allowlists - Isolated network namespace per agent - Bring your own auth (API keys, OAuth, JWTs) Same sandboxing that protects billions of Chrome tabs. Now protecting your agents.

Built-in orchestration. No glue code. - Webhooks for external events - Multiplayer: multiple clients observe one agent - Agent-to-agent delegation - Durable workflows with retries - Queues for backpressure - Persistent SQLite for structured state Everything agents need to coordinate, out of the box.

agentOS is open source and runs anywhere: Your laptop. Railway. Vercel. Kubernetes. ECS. Lambda. Cloud Run. A Raspberry Pi. One npm package. No vendor lock-in. Same API everywhere.

GitHub:

Get started:

How is this different from Anthropic's srt?

these are genuinely insane benchmarks

Insane stuff

It's been three days, time to upgrade everything again. 🤣

- how running code in a sb can access dependencies (npm) - how does the workflow persist (where ? how to plug the infra ?) I have so many questions. It looks like it's so many things at once it's hard to to figure this out. I can run sb code on same machine, it can run sb code in sb provider, it can persist sb, it can run full harness and act as a server....... 🤯 perhaps, what are the things we can't do with this ? maybe this would be easier to take it this way

The "OS for agents" framing is doing a lot of heavy lifting here. Isn't the real unlock the file system abstraction, not the OS layer? 6ms coldstarts are cool but who's actually bottlenecked by coldstarts vs context switching costs?

Great stuff

I need this now!

Nice

The model war is mostly over. The infra war is just starting. Claude Code, Codex, and OpenCode all converging on the same execution layer is the platform consolidation story nobody's talking about yet.

legends

wow

I'm rooting for rivet to take off even when using cloudflare for production use cases.

beautiful launch, can't wait to try!!

Beautiful

Very slick!

Any way for someone with a python stack to take advantage of this?

@bikramkgupta check this out.

Cool initiative!

Looks cool. Wish it ran in Cloudflare. Get them to support V8.

good luck w the SEO on that name!

🔥🔥

Amazing 🚀

This is gonna be one hell of you ride 💪 Can't wait to try

@grok come up with 10 product ideas 💡 using AgentOS

this is cool 🫡

@MoritzW42 Can be interesting to generate roolouts

1this is very interesting

really love your new agentOS API + impl. so clean, from spawning vms to listening to sessions + prompting the combo w/ actors make it so valuable would be so cool to launch you guys on if that helps.

this is another crazy one @iamlukethedev still early

6ms coldstarts changes the dispatching math entirely. right now most agent setups run one session at a time because startup cost kills parallelism. at 131MB per agent you could run dozens concurrently on a single machine and let them specialize rather than asking one agent to do everything sequentially that's when agent fleets start making economic sense

this is really cool! We need our agents to do computer use and potentially host a server to test its changes. Is that possible with agentOS?

6ms coldstarts for agent sandboxing is no joke. Running untrusted agent code without destroying your infra is the hardest unsolved problem right now.

okay pack it in, it's over, this is clearly the final form factor for agent runtimes