Google Cloud, part of Alphabet Inc. (Nasdaq: GOOGL), has introduced Agent Compute Engine, a purpose-built service designed to give AI agents isolated computing environments while helping enterprises govern their actions across cloud and hybrid environments. The service is scheduled for preview in the coming months.
The product assigns each agent a hardware-isolated microVM rather than a conventional shared container or long-running virtual machine. Google Cloud said the design is intended for agents that run code, use tools, access browsers or orchestrate workflows, then pause while awaiting a model response, a person or an outside system.
For infrastructure teams, the practical issue is how to give agents the filesystem, runtimes and network access needed to complete tasks without leaving an always-on machine allocated to every idle session. Google Cloud said Agent Compute Engine can pause an idle microVM through checkpoint and restore, then resume it when the next agent step begins. The company also said copy-on-write memory sharing lets multiple agents use one operating-system template while retaining separate execution environments.
Security is central to the service’s enterprise case. Google Cloud said Agent Compute Engine uses its Titanium infrastructure processing unit and a Cloud Hypervisor boundary to isolate each guest kernel. Each microVM receives its own identity, while service credentials are injected outside the microVM’s egress path. That approach is meant to reduce the chance that untrusted code running inside an agent sandbox can directly access shared credentials.
The service also provides each microVM with a Hyperdisk for persistent workspaces and agent context, according to the company. Google Cloud said it integrates the runtime with its identity and cloud services so organizations can apply the same access model and auditing practices to agent connectivity and data access.
For teams that need more control or portability, Google Cloud is also positioning Agent Substrate on Google Kubernetes Engine as an open-source, Kubernetes-native execution runtime. The company said Agent Substrate can use microVMs or gVisor for isolation, return an idle machine to a shared pool and restore agent state when it is needed. Google Cloud reported benchmark validation of 500 activations per second and less than 500 milliseconds of suspend-and-resume latency while running 200,000 agents; those figures are company results.
Alongside the compute service, Google Cloud is expanding Gemini for business governance to cover Agent Runtime, Cloud Run, GKE and Agent Compute Engine, with planned cross-cloud and managed-endpoint coverage. The controls include an Agent Gateway for policy enforcement, an Agent Registry for approved agents and tools, agent identities based on SPIFFE, and OpenTelemetry-compatible observability. Forthcoming features include token budgets, model-routing policies and extensible policy bindings, the company said.
The immediate benefit for platform and security teams is a more integrated way to isolate execution, manage identities and observe agent activity without assembling every runtime control independently. Availability remains limited to a forthcoming preview, and enterprises will still need to define their own access policies and review processes. But the combination targets a growing operational problem: allowing agents to act on enterprise systems while retaining boundaries, audit trails and resource controls.