Why Keep an Agent VM Running While the Model Is Thinking?
Your agent’s shell command finished seconds ago, but the model is still thinking and the next tool call needs those files. Do you pay for the sandbox the whole time it waits?
- Agent Compute Engine preserves working state while reclaiming idle sandbox resources.
- The model-calling harness runs separately from the code-execution environment.
- Customer preview is still ahead; Agent Substrate provides a separate Kubernetes path.
The model can think while its execution sandbox sleeps
An agent waiting for its next model response may have no code left to execute. Google’s Agent Compute Engine design uses that gap: the sandbox signals when no commands or background tasks remain, allowing the host to checkpoint it, reclaim resources and restore it for a later tool call.
The harness stays outside the sandbox and decides what to do next. A separate microVM runs the requested commands and browser work. Suspending that execution environment does not require restarting the entire agent from its original task.
A restored workspace still needs a valid next action
Persistence matters when work spans multiple sessions. Google’s October 22 announcement describes a Hyperdisk for each microVM to retain workspaces and agent context. Reclaiming idle compute is compatible with retaining work.
That infrastructure distinction has an operational consequence: “ready to execute again” does not mean “authorized to continue.” If a workflow stopped because a person must choose a deployment target, restoring its machine cannot supply that decision. The harness still needs to determine the next permitted action.
Resuming a machine does not grant unrestricted access
Google keeps long-lived credentials outside the execution sandbox and brokers external connections. The Agent Gateway documentation describes the complementary controls: an agent identity, a registry of approved destinations, and policies governing access and content.
Isolation separates execution environments. Gateway policy controls what an environment can reach. A retained workspace therefore should not be treated as permission to perform any action available through its tools.
Availability also varies by feature and runtime. The current gateway documentation marks expanded runtime coverage and token-management policies as forthcoming. Their inclusion in the broader announcement does not make them immediately available everywhere.
Managed preview is ahead; Kubernetes has a separate route
As of October 26, the announced customer window for Agent Compute Engine remains a preview in the coming months. Google already uses it for Gemini, but that is different from general customer availability.
Teams operating Kubernetes can separately evaluate Agent Substrate, whose GKE availability was announced on September 15. Google describes it as an open-source runtime portable across Kubernetes infrastructure, with suspension of idle execution. It is not the managed Compute Engine product under another name.
For an existing deployment, measure active tool execution separately from model and human waiting time. Long waits combined with persistent workspaces make suspension and restoration worth evaluating alongside VM sizing.