Google Agent Compute Engine: a dedicated compute environment for AI agents

What it's about

Google has unveiled the Agent Compute Engine, a compute service built specifically for AI agents. Agents are AI programs that complete tasks on their own in multiple steps, such as running code or using tools. To do that, they need an environment to compute in. According to Google, the service already runs behind Gemini and should become available to customers in the coming months as a preview, meaning an early test version. The report comes from The Decoder and contains few details.

How it works

Each agent gets its own microVM, a lightweight virtual machine that works like a separate mini computer. It is secured by hardware isolation through Google's Titanium chip. This is meant to keep one agent from accessing another's data. According to Google, the microVMs start quickly, pause automatically when idle and share memory to save resources. That makes sense because agents often spend a long time waiting, for example for a model's reply.

Agent Substrate as an open-source option

Google is also offering Agent Substrate, an open-source alternative based on Kubernetes, a widely used piece of software for managing containers. According to Google, it reaches up to ten times the compute density, meaning significantly more agents per machine. In a benchmark, 200,000 agents ran with latency under 500 milliseconds. These are vendor claims, and the source gives no details on the test setup.

Governance and cost control

Google is extending its governance features (rules and controls over what agents are allowed to do) to more environments: Cloud Run, GKE (Google's Kubernetes service) and local endpoints. Also new:

• Token management for cost control
• Model routing, meaning requests are forwarded to suitable models
• Authorization through third parties such as Ping Identity and SailPoint

Takeaway

As the number of agents grows, secure, affordable runtime environments become an infrastructure question. Google is positioning itself with its own hardware isolation and an open variant. Pricing, the exact launch date and independent measurements of the performance claims remain unclear. This matters mainly for companies and developers who want to run many agents, and less for end users.