Google announced Agent Compute Engine for running AI agents. Its customer preview is described as arriving in the coming months. It should not be presented as a generally available service that every customer can deploy today. Google Cloud official announcement
The announcement emphasizes microVMs with their own kernels, session pause and restore, agent identities, and short-lived credentials. It treats the computer that executes code or drives a browser as a distinct infrastructure layer.
On this page
An agent needs a workspace as well as a model
A chatbot may produce a response directly, while a tool-using agent can create files, execute code, and inspect the results before continuing. Model serving and the workspace where actions run have different responsibilities. Improving reasoning quality does not resolve workspace isolation or authorization. The execution computer therefore deserves its own operating requirements and security review.
A browser task can leave login cookies, downloaded files, and temporary documents behind. If a workspace is reused, confirm that one user’s state cannot affect the next user’s task. Session cleanup and data separation are requirements in their own right, alongside the amount of CPU or memory available to the agent.
Why microVMs and pause-and-resume matter
A microVM is a virtual machine with a deliberately limited, lightweight design. Unlike containers that commonly share the host kernel, a VM can establish a boundary with a separate guest kernel. That does not remove every risk. Shared files, network access, and management interfaces still determine what can cross the boundary and what the workload can reach.
Agents may alternate between brief tool execution and waiting for model responses or human approval. Pausing and restoring a waiting workspace changes how idle resources can be managed. Restored execution state must not automatically restore authorization that has expired. Treat the saved computer state and the current permission state as separate concerns, especially for tasks lasting across long approval delays.
Map interruption and resumption to the business workflow
For a hypothetical reporting agent, reading source material and saving a draft might be allowed automatically while external delivery requires approval. If the task resumes the following day, recheck the destination, selected document, and current permissions. The ability to restore a session does not establish that every business action in that session remains authorized.
Document files left after forced termination, changes already committed to external systems, and actions that would be duplicated by a retry. Payments or message delivery need particular attention to duplicate prevention. When the preview becomes available, these concrete scenarios provide a more useful comparison than an isolated demonstration of a fast startup.
Requirements you can define before release
- Specify network destinations and filesystem access per agent.
- Define credential and workspace cleanup at session termination.
- Document persistence, resume, and forced-stop requirements for long-running jobs.
For now, establish evaluation requirements rather than replacing an existing runtime. Compare the actual preview offering against them when access conditions are published.
Announcement reference date: 2026-10-22 · Checked: October 24, 2026
