Micro-VMs for AI agents that cold-start in under 60 ms

What it does

CubeSandbox is a Rust-based sandbox service built on RustVMM and KVM that spins up fully isolated, E2B-compatible execution environments for AI agents. Each sandbox runs its own guest OS kernel rather than sharing the host kernel like Docker, and the system supports both single-node and clustered deployments. It targets the same use case as the E2B closed-source sandbox, but claims sub-60-millisecond cold starts and less than 5 MB of memory overhead per instance.

The interesting bit

The project achieves its speed through resource pool pre-provisioning and a Copy-on-Write snapshot engine called CubeCoW, which lets it fork and roll back running sandboxes at millisecond granularity. That means you get the isolation of a traditional VM—dedicated kernel, eBPF-enforced network policies via CubeVS, and no container escape risks—with a footprint and boot time that rival containers.

Key highlights

• Drop-in E2B SDK compatibility: change a URL and existing agent code runs unmodified.
• Sub-60 ms cold starts on bare metal (P99 under 150 ms at 50 concurrent creations).
• Per-instance memory overhead below 5 MB for configurations up to 32 GB, enabling thousands of agents per node.
• Event-level snapshots and instant cloning via the CubeCoW engine.
• Production-hardened: Tencent Cloud runs it at scale, and the stack is Apache 2.0 licensed.

Caveats

• Requires x86_64 Linux with KVM; WSL2 and ordinary cloud VMs (via PVM) are supported workarounds, but the README never mentions macOS or ARM.
• Memory overhead claims are benchmarked against sandboxes sized at 32 GB or less; larger specs may cost marginally more.
• The nested-QEMU development environment is explicitly flagged as low-performance and not recommended for real workloads.

Verdict

Teams running untrusted LLM-generated code at high density should look here, especially if they are already paying for E2B. If you are not on x86_64 Linux or unwilling to use KVM, look elsewhere.