Google designs persistent agent computers, with customer preview still ahead

The proposed service gives each agent a hardware-isolated microVM, while persistent state lets files and installed tools survive across sessions. Google says this architecture already supports Gemini agents internally; the announced customer preview is still expected in the coming months. [1][2]

The design treats waiting time as a major cost problem. When an agent has no active commands or background work, Google checkpoints its state and releases active host resources, then resumes it for the next tool call. Snapshot, fork and rollback operations support parallel experiments without rebuilding the environment. These are Google’s architecture claims, not externally measured service-level guarantees. [1]

Security is also part of the runtime design: an off-node gateway checks identity and policy and inserts backend credentials into outbound requests, keeping those credentials out of the sandbox. That can narrow one route for credential theft, but it does not prove that every permitted agent action is safe. Public pricing and generally available customer service are not established by these announcements. [1][2]