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

Google on October 22 detailed its Agent Compute Engine, which gives each AI agent session a hardware-isolated, single-tenant microVM that is checkpointed during model inference, cloned copy-on-write from warm snapshots, and stripped of a network card, with preview promised "in the coming months." The post omitted startup, resume, density and cost figures and did not name the hypervisor, so a developer rebuilt the pattern on Firecracker in roughly 2,200 lines of Go and benchmarked it on two machines. On a Mac mini under nested virtualization, twelve concurrent forks never became ready, exposing how sensitive the approach is to the host environment. The same day, the Agentic AI Foundation added four projects including Alibaba's OpenSandbox and Meta-PyTorch and Hugging Face's OpenEnv, signaling consolidation around execution environments for agents. What to watch: Google's first published latency and density numbers for Agent Compute Engine, and whether open Firecracker-based builds close the gap.