Google Cloud has announced Agent Compute Engine, a compute service designed around the lifecycle of interactive AI agents. Its October 22 account says the technology already supports Gemini work tasks, while customer preview access is planned in the coming months. That timing is important for teams considering it for a near-term deployment.
Agent sessions pause as well as execute
In Google Cloud’s announcement, agents use microVMs that can start, pause and resume. The service is designed to checkpoint idle work while an agent waits for a person or an external system, then restore the session when it can continue.
Google describes hardware-isolated execution, individual identities and short-lived credentials. Each microVM also has persistent storage for its workspace and context. These are the company’s architecture claims; their suitability for a particular workload depends on the actual service configuration and access available at preview.
Kubernetes teams have a related runtime path
The announcement also describes Agent Substrate, an open-source runtime for Kubernetes. It snapshots idle agents and returns compute to a pool, with a design intended to support customization and portability.
Google reports performance figures from its own benchmark validation. Those results belong to the stated setup and should not be treated as guaranteed density or resume latency for every cluster. The relevant evaluation includes the agent’s tools, persistent state and network requirements.
Governance extensions have separate milestones
Google describes identity, a registry, a gateway and observability as parts of its broader agent-governance approach. Cross-cloud enforcement, local-agent controls and additional token and routing capabilities include forthcoming preview milestones.
A practical planning exercise is to map each required control to its release status and operating responsibility. The announcement brings compute and governance into one discussion, but it does not make every listed capability generally available at once. Teams can prepare a representative workload and evaluate it against the actual preview rather than infer production readiness from the launch language.