Google has unveiled Agent Compute Engine, a new Google Cloud service that hands each AI agent its own hardware-isolated micro virtual machine. The Google Agent Compute Engine already runs Gemini’s own agents behind the scenes, and Google says business customers will get it in preview in the coming months. Alongside it, the company released an open-source runtime for Kubernetes and a batch of new governance controls for agent fleets.
What Google Agent Compute Engine Actually Does
Modern AI agents do more than chat. They write and run code, open browsers, call tools and work through long, multi-step tasks, so each one needs a real computer with a file system, language runtimes and network access.
Google’s answer is a “microVM” for every agent. Each one gets its own guest kernel and root access, sits behind a hardware-isolated hypervisor boundary, and runs on Google’s Titanium offload chip for an extra layer of protection.
The clever part is how the service handles idle time. When an agent finishes a tool call and starts “thinking,” its microVM can pause through a fast checkpoint and resume the moment it’s needed again, so the machine isn’t wasted while the model works.
Why regular servers fall short
Google explains the trade-off in a technical deep dive. Traditional VMs are strongly isolated but take seconds or minutes to start, pause and resume, while containers are fast but share the host’s kernel, which isn’t safe enough for untrusted code an agent might run.
Agent Compute Engine aims to combine both: the isolation of a VM with the speed and density of a container. It also shares memory between agents built from the same template, which cuts the cost of running each one.
Key Features at a Glance
| Feature | What it means for agents | Status |
|---|---|---|
| Agent Compute Engine | A hardware-isolated microVM per agent, with its own persistent disk | Powers Gemini now; customer preview in coming months |
| Agent Substrate | Open-source Kubernetes runtime that snapshots idle agents and frees the machine | Available to try on GKE |
| Token and cost governance | Track token spend per agent and set budget guardrails | Preview this quarter |
| Model routing | Send each request to the most suitable model to control cost | Preview this quarter |
| Third-party identity support | Delegated user sign-in through Ping Identity, SailPoint and Saviynt | Preview this quarter |
Agent Substrate Brings the Same Idea to Kubernetes
Not every team wants a fully managed service. For companies already standardized on Kubernetes, Google released Agent Substrate, an open-source runtime for Google Kubernetes Engine that follows the same agent-first design.
Google says Agent Substrate can deliver up to 10 times higher compute density than traditional compute while keeping full kernel isolation. In a benchmark with 200,000 agents, it reported sustaining 500 activations per second with suspend and resume taking under 500 milliseconds.
Because it is open source, it runs on any Kubernetes cluster and works with popular agent harnesses, including Claude Code. That fits a wider industry push toward safe agent sandbox environments for code that models write on the fly.
New Governance Tools for Google Cloud AI Agents
The second half of the announcement is about control. Google is extending the agent governance features in its Gemini for business platform so they cover agents wherever they run, not just inside Google Cloud.
- Agent Gateway acts like air traffic control for traffic between users, agents and tools, applying data protection and attack protection policies.
- Agent Registry keeps a central list of approved agents, tools, MCP servers, skills and models.
- Agent Identity gives each agent a unique, standards-based ID for sign-in, access control and audits.
- Agent observability emits standard OpenTelemetry metrics, logs and traces for every model call and tool use.
Google also plans to bring governance to coding agents such as Claude Code and its own Antigravity CLI on managed laptops, using an OS-level sandbox. Similar thinking shows up in recent tools for agent tracing data and enterprise agent governance controls from other vendors.
The race in AI agents is shifting from smarter models to safer plumbing. Whoever runs agents cheaply and securely at scale gets a real edge.
What This Means for You
If you only use Gemini as a consumer, you won’t see a new button. You may notice agents that start faster and keep their work between sessions, since this is the system Google uses to give each Gemini agent its own computer.
If your company builds or runs agents, here are a few practical steps:
- Register interest in the Agent Compute Engine preview if you need isolated sandboxes for agents that run untrusted code.
- Try Agent Substrate on a test GKE cluster if your team prefers open source or multi-cloud setups.
- Plan for per-agent identity and token budgets now, because governance is quickly becoming a buying requirement.
You can follow more of Google’s recent AI moves, such as the Gemini 4 launch, and other agent stories on our AI news page.
Sources
Primary source: Google Cloud’s Agent Compute Engine announcement. Technical details from Google Cloud’s Agent Compute Engine deep dive. Additional reporting from The Decoder.
Conclusion
Agent Compute Engine shows how seriously Google is taking the cost and security problems of running millions of AI agents at once. By giving each agent its own isolated microVM, pausing idle ones and wrapping everything in identity and budget controls, Google is betting that infrastructure will decide who wins the agent era. The managed service is still months from customer preview, but the open-source Agent Substrate lets teams start experimenting today. For businesses planning agent rollouts, the message is clear: isolation and governance now matter as much as the model itself.
