Google Described Its Agent Sandbox. I Built One with Firecracker and Measured It

A developer built an open-source implementation of Google's Agent Compute Engine sandbox using Firecracker microvirtual machines to measure its performance and feasibility. The implementation achieved 25ms snapshot restores (26x faster than cold boots) and demonstrated that 64 idle VM forks consume only 3 MiB of memory each on bare metal, though performance degraded significantly under nested virtualization. The work highlights how AI agents require stronger isolation than containers provide since they execute untrusted code, making hardware-isolated microvMs a practical solution for secure agent sandboxing.