An in-process OS that isolates AI agents with sub-10-ms cold starts

What it does

agentOS is an in-process operating system kernel that spins up isolated environments for AI coding agents inside your existing application. Instead of booting a Linux VM or pulling a container, it mounts a virtual filesystem, process table, and network stack on top of WebAssembly utilities and V8 isolates. Agents like Pi run inside this lightweight VM, executing shell commands and Node.js scripts while remaining sandboxed from the host.

The interesting bit

The architecture is a JavaScript kernel simulating an OS inside your process, where POSIX tools are WASM binaries and agent logic runs in V8 isolates—the same isolation model browsers use. This lets agents invoke host-defined JavaScript functions as if they were local CLI commands, eliminating network hops between your backend and the agent.

Key highlights

• Cold starts measured at 4.8 ms median and 6.1 ms p99 against E2B’s 440 ms and 3,150 ms, according to the project’s March 2026 benchmarks.
• Memory footprint of roughly 22 MB for a shell command and ~131 MB for a full coding agent with filesystem, compared to sandbox baselines starting around 1,024 MB.
• Deny-by-default permissions for filesystem, network, and process access, plus programmatic network proxying and per-agent resource limits.
• Built-in orchestration primitives: multiplayer sessions, agent-to-agent delegation, durable workflows with retries, cron, webhooks, and queues.
• A registry of pre-built WASM packages including coreutils, ripgrep, jq, and curl; some staples like git and make are still listed as planned.

Caveats

• Most named agent integrations—Claude Code, Codex, Amp, and OpenCode—are currently marked “coming soon,” with only Pi demonstrated in the quickstart.
• Benchmark cost comparisons assume a self-hosted deployment at 70% utilization; managed-cloud savings will differ.
• The registry still lists foundational developer tools like git and make as planned rather than available.

Verdict

Teams running high-frequency or embedded coding agents should look here, especially if millisecond latency and tight backend integration matter more than running native binaries or graphical browsers. If you need a full Linux environment for those heavier workloads, the sandbox extension exists, but you might just stick with a traditional sandbox instead.