Zeroboot bets AI agents get real KVM isolation at sub-millisecond fork cost

Zeroboot is an open-source infrastructure project, and early-stage startup, betting that AI agents will execute untrusted code in real virtual machines if doing so costs as little as a function call. Its Rust engine forks Firecracker microVMs from a copy-on-write memory snapshot: p50 spawn latency of 0.79ms and roughly 265KB of memory per sandbox, each a separate KVM VM with hardware-enforced memory isolation.

The mechanism is the bet: boot one template once, preload the runtime, snapshot memory and CPU state, then mmap(MAP_PRIVATE) to fork thousands of VMs. That makes speculative parallel execution practical — fork ten VMs, try ten approaches, keep the best — and turns code execution from an infrastructure decision into a cheap tool call, which is what agent frameworks such as LangChain, CrewAI and OpenAI Assistants need.

Show HN on 2026-03-17 drew 311 points and 72 comments, with engineers comparing the approach to CodeSandbox's microVM cloning and Cloud Hypervisor's userfaultfd work. The author answered that single-node density is the near-term point (1,000 concurrent sandboxes on one roughly $50 box) and that a managed API was already live, with plans for multi-region deployment, custom templates and usage-based pricing.

As of 2026-09-05 the repository shows 2.4k stars and 108 forks, and the README still labels the project a working prototype: the fork primitive, benchmarks and API are real but not production-hardened, with documented limits of one vCPU per fork, no in-fork networking, and a ~15s full re-snapshot on template updates.