Google explains the infrastructure behind persistent AI agent computers

Agent Compute Engine pairs isolated microVMs with rapid pause and resume, aiming to reduce the cost of waiting agents.

Google has published a technical account of Agent Compute Engine, the infrastructure it built to give AI agents persistent computing environments. The October 22 post describes a challenge beyond model inference: agents also need files, browsers and tools, yet spend substantial time waiting between actions.

Google says each session receives a dedicated hardware-isolated microVM. The design keeps long-lived credentials outside the guest and brokers network access through a separate mediation path. These are architectural descriptions from the service’s developer, not an independent security certification.

Preserving work without reserving an idle computer

The post explains how application-aware idle detection can checkpoint an environment after its commands finish and restore it for the next action. Shared base images and memory pages reduce duplication, while the user’s workspace persists separately from disposable scratch data.

That distinction is useful for understanding agent infrastructure. An agent may need yesterday’s files even when it has done no useful computation for hours. Keeping state and continuously reserving active resources are different requirements.

Sharing the base, separating the work

Google describes pre-warmed snapshot images whose unchanged filesystem and memory pages can be shared across sessions. Copy-on-write allocates private pages when an agent modifies state. The design separates disposable scratch data, read-only shared materials and durable user workspaces.

That is more specific than simply packing more agents onto a server. It treats a waiting session, shared runtime and saved customer work as different resource problems. The security design likewise separates long-lived credentials from the sandbox and mediates outbound requests. The architecture’s promise is to retain useful work while reducing duplicate resources; actual cost and recovery behaviour still depend on the workload.

Sources

- [Inside Agent Compute Engine: Building the compute foundation for the agentic era](https://cloud.google.com/blog/topics/ai-infrastructure/agent-compute-engine-deep-dive/) — Google Cloud