Google Cloud unveils Agent Compute Engine for isolating AI agents

Google Cloud has announced Agent Compute Engine, a compute service built to run each AI agent inside its own microVM, a lightweight virtual machine that isolates a single workload from everything else on a server. The company says the service already runs its own Gemini agents and will open to a wider preview in the coming months.

The core idea is isolation. Instead of multiple AI agents sharing a server process or container, Agent Compute Engine gives each one hardware-level separation, meaning the virtualization is backed by the physical machine rather than just software boundaries within an operating system. Google says these microVMs carry zero in-process credentials, so the agent itself never holds API keys, passwords or access tokens inside its own running environment.

That distinction matters for how AI agents are typically built today. Agents that can browse the web, call APIs or execute code often need credentials to do that work, and those credentials have usually lived inside the same process as the agent’s logic. If an agent is compromised, hijacked through a malicious prompt, or simply misconfigured, credentials stored that way can be exposed. Google’s announcement frames Agent Compute Engine as a way to separate the credential from the compute, though the company has not detailed exactly how credentials are supplied to an agent’s actions if not held in process.

Google Cloud has not published pricing, regional availability or which customers will get early access once the preview opens. It also has not said whether Agent Compute Engine will support agents built on models other than Gemini, Google’s own AI model family, or whether it is designed to work with third-party agent frameworks. The announcement states only that it already runs Gemini agents internally and that broader preview access is coming within months, without a specific date.

What’s not yet known

The announcement leaves open several practical questions that will matter to any team evaluating the service. It is not clear what the performance or cost overhead of a dedicated microVM per agent looks like compared with running agents in shared containers, a tradeoff that often shapes whether security-focused infrastructure gets adopted at scale. It is also not clear how the service handles communication between an agent and the credentials or secrets it needs to act on a user’s behalf, since the whole premise rests on those credentials staying outside the agent’s own runtime.

Google Cloud has not said when preview access will open beyond “coming months,” nor shared documentation on setup or supported agent frameworks. Teams building or securing AI agents should watch for Google’s preview documentation and pricing details once they appear, since those will determine whether Agent Compute Engine is a practical option or a Gemini-specific showcase for now.