Google Agent Compute Engine: MicroVMs That Pause Idle AI Agents

Google Cloud announced Agent Compute Engine, a runtime that gives every AI agent its own hardware-isolated microVM and pauses it while the agent waits. It already runs Google's own Gemini agents. You can't use it yet: customer preview is "in the coming months," and Google published no pricing. The real news for operators is the problem it targets. Long-running agents spend most of their life idle, and you pay for that idle time.

What actually happened

Per Google Cloud's October 22 announcement:

• Isolation: each agent runs in a Cloud Hypervisor microVM with its own guest kernel, on Google's Titanium hardware. Each microVM gets its own verified identity and short-lived credentials. Service credentials are injected at egress, outside the microVM, so the agent never holds them.
• Pause and resume: when an agent finishes a tool call and starts "thinking," the microVM signals idle and the host checkpoints it. It resumes from a snapshot when a human or outside system answers. Copy-on-write memory lets many agents share one OS template.
• State: each microVM gets its own Hyperdisk, so workspace and context persist across long sessions.
• No numbers for the product itself. The benchmark figures in the post (500 activations per second, under 500 ms suspend and resume across 200,000 agents) belong to Agent Substrate, a separate open-source project on GKE.

Google also expanded Agent Gateway with token budget guardrails, spend attribution and model routing policies, all in preview "this quarter," plus governance for Claude Code and Antigravity CLI on managed laptops, "in preview soon."

Why it matters for your business

Idle time is the hidden line on your agent bill. An agent that waits two hours for a manager to approve a refund still holds a container, memory and often a warm sandbox. Pausing that is the right fix. Until a runtime does it for you, design your agents to save state and exit while they wait, then restart from a queue event. We do this today on plain serverless functions.

Credentials outside the box is the pattern to copy. Google keeps service keys out of the microVM. You can do the same on any platform: put a proxy between the agent and your APIs, and give the agent nothing it could leak.

Budget guardrails are coming everywhere. Don't wait for them. Set a hard per-agent token cap and a spend alert at your model provider now.

Watch the lock-in. Snapshots, identity and gateway policy all tie to Google Cloud. Keep agent logic portable and treat the runtime as replaceable.

Sources: Google Cloud blog, InfoWorld.