An engineer made an open copy of the Google agent sandbox

On 22 October, Google showed Agent Compute Engine, a cloud system that gives each AI agent one sandbox. An engineer then made a copy of the system with Firecracker, an open source tool for small virtual computers. The engineer measured it. On a physical server, the time for a start from a snapshot is 25 ms. That is 26 times as fast as a cold start. The Google post gave no speed numbers.

How it works

Each agent session operates in one microVM, a small virtual computer with a different kernel from the host. Many microVMs start from one snapshot. They use the same memory pages until a VM changes a page. The guest has no network card. It can only connect to a proxy on the host that holds the credentials.

The numbers

• 64 unused copies of one VM use about 3 MiB of host memory each.
• 64 copies that start at the same time all start in about 150 ms.
• On a Mac mini with nested virtualization, 12 copies did not start.

What is not known

Google gave no numbers for startup, density or cost. It did not name its hypervisor. The preview is planned for the coming months, Google says. The test is of the copy, not of the Google system.

Insight: Sandboxes are easy to copy, so the credential gateway matters

Google gave no numbers, but one engineer made the same system with open parts. On 22 October the Agentic AI Foundation also added 2 sandbox projects, from Alibaba and from Meta-PyTorch with Hugging Face. Akamai and AWS also have sandboxes. A sealed sandbox seems a standard part, and control seems to move to who holds the credentials. Find out which gateway your agents must use.