2026年3月18日InfrastructureOpen SourceTool

Zeroboot: Sub-Millisecond VM Sandboxes for AI Agents

Zeroboot is a new open-source project that creates lightweight virtual machine sandboxes optimized for AI agent code execution, achieving spawn latency of 0.79ms (p50) with just ~265KB memory per sandbox.

Created on March 15, 2026, Zeroboot has quickly gained attention — 376 GitHub stars and 200 points on Hacker News. It uses Firecracker (a lightweight hypervisor) combined with memory-mapped copy-on-write techniques to fork real KVM virtual machines near-instantly. Each forked environment provides hardware-enforced memory isolation, not just container-level separation.

The performance numbers are striking: fork + execute cycle takes ~8ms for Python, and 1,000 concurrent forks complete in 815ms. The project provides SDKs for Python and TypeScript/Node.js, plus an HTTP API for remote code execution. It creates a template VM once, then forks with restored CPU state in under one millisecond.

This directly addresses a pain point in agentic systems where agents need to execute untrusted code safely and at scale. Written in Rust, Apache 2.0 licensed.

https://github.com/adammiribyan/zeroboot
← 上一篇
discli:让AI代理从终端完全访问 Discord 的命令行工具
下一篇 →
GPT-5.4 Mini 与 Nano:OpenAI 为子智能体时代打造的模型
← 返回所有文章

评论

加载中...
>_