@motatoeshq: How we're scaling http://opencomputer.dev to 1M sandboxes
Summary
OpenComputer offers long-running, persistent cloud VMs for AI agents, enabling stateful, always-on compute with dynamic resizing, as an alternative to ephemeral sandboxes.
View Cached Full Text
Cached at: 06/08/26, 07:31 PM
How we’re scaling https://t.co/TQuSY4M9Bp to 1M sandboxes https://t.co/7DoojRtAsa
OpenComputer – Long-running cloud infrastructure for AI agents
Source: https://opencomputer.dev/
Beyond sandboxes.
Today, agents use sandboxes to run untrusted code. Disposable computers that spin up, do a task, and disappear. But agents are getting more ambitious. They need a whole computer at their disposal - always on, always persistent, always ready.
Sandboxes are for throwaway tasks. Agents need something that sticks around.
It’s time to give your agents a real computer.
Every OpenComputer is a real machine — a full filesystem, full OS access, and persistent state. It stays always on, ready when you need it. No timeouts, no teardowns. Your computer is just there.
Resize memory and CPU at runtime to match your workload. When you don’t need it, hibernate and wake it up at any point — your state is exactly where you left it. It scales to thousands in the cloud, and you only pay for what you use.
Built for B2B agent platforms.
If you’re building the next Lovable, Devin, or Bolt, your users don’t just need a sandbox that runs a script and dies. They need a computer that remembers what it installed yesterday, keeps their files between sessions, and picks up exactly where it left off. Sandboxes give you isolation. OpenComputer gives you isolationandpersistence.
Ephemeral sandboxes are stateless - every session starts from scratch. OpenComputer VMs are persistent - they stay on until you explicitly stop or delete them, so state survives across sessions without any extra work.
No more re-installing node_modules from scratch because the container timed out. Your VM stays alive as long as you need it. Need more CPU mid-session? Resize on the fly without restarting.
Similar Articles
@jhleath: https://x.com/jhleath/status/2065408690992148698
The author explains how they built a compute platform capable of launching millions of sandboxes per second in constant time, focusing on decoupled scheduling and capacity aggregation using Cassandra and S3.
@latentspacepod: Daytona’s Agent-Native Compute: 60ms sandboxes, 50K startups in 75 sec, 850K daily runs, RL/evals, CLI > MCP, & the end…
Daytona's CEO Ivan Burazin discusses their agent-native compute platform offering 60ms sandboxes, stateful snapshots, and support for RL/evals, marking a shift from local development to cloud-based agent infrastructure.
@zongheng_yang: Sandboxes are all the rage (Modal, E2B, AWS, ..). Most AI teams pay a >4x markup to run sandboxes on someone else's mac…
SkyPilot Sandboxes allows AI teams to run sandboxes on their own clusters, offering 4-10x cost savings compared to Modal with sub-second launches and warm pools.
@claudeai: Live from Code with Claude London: we're launching self-hosted sandboxes (public beta) and MCP tunnels (research previe…
Anthropic launches self-hosted sandboxes (public beta) and MCP tunnels (research preview) in Claude Managed Agents, enabling agents to run within the user's own perimeter with default security controls.
How We Built Secure, Scalable Agent Sandbox Infrastructure (8 minute read)
Browser Use describes two patterns for isolating AI agents that execute code: isolating the tool vs isolating the agent. They implemented the agent isolation pattern using Unikraft micro-VMs on AWS, achieving secure, scalable, and disposable sandboxes.