Google introduced Agent Compute Engine, a service for running AI agents in hardware-isolated microVMs, at AGNTCon+MCPCon on Oct. 22, 2026, The New Stack reported. SiliconANGLE also reported the introduction. The New Stack says the service began as internal infrastructure for Gemini, Antigravity and Google's model training pipelines. A preview for Google Cloud users was planned for later in 2026, according to The New Stack; SiliconANGLE described a rollout in the coming months.
Credentials stay outside the sandbox
According to The New Stack, each agent session runs in a single-tenant, memory-safe microVM designed to hold no secrets in memory or on disk. The microVM has no standard network interface; traffic goes through a proxy outside the sandbox. When an agent needs to access a service such as Google Drive, Agent Gateway checks the user's delegation and the sandbox's identity before adding a credential outside the sandbox. These are descriptions of the system's design, not proof of its resistance to attacks.
Checkpointing between calls
Google engineers Rohit Jnagal and Ramanathan Venkatapathy described a checkpointing mechanism in The New Stack's account: after a tool call, the host checkpoints the microVM, which then uses no active resources until the next call. SiliconANGLE reports that Google says idle sandboxes shut down and reactivate when needed. It also quotes the engineers describing cloning, snapshot/forking, rewind and rollback for agent harnesses. Neither account gives a measured cost saving for Agent Compute Engine.
Comments