Google Cloud introduced Agent Compute Engine on October 22, 2026, at the AGNTCon and MCPCon events. The service is designed to host AI agents within hardware-isolated microVMs, each featuring a dedicated guest kernel and root access.
The platform utilizes Titanium IPU technology for security and supports features such as rapid startup, pausing, and resuming workloads. It also employs memory sharing, where agents share a base memory until modifications occur, triggering the creation of a private copy.
Google reports that this architecture achieves 10x density, allowing for a higher number of agents on existing hardware. The service is currently integrated into the Gemini Enterprise Agent Platform and is already used to power Google's internal Gemini agents.
While the technology is operational internally, public access is currently limited. Google has announced plans to launch a preview phase for customers soon, though a specific date for the public release has not been disclosed.
