Running large numbers of AI agents requires both isolation and efficient use of computing resources. Google’s Agent Compute Engine is designed to give each agent its own virtual machine while reclaiming machine capacity when that agent is idle.
Each agent gets an isolated environment
Google says every agent runs inside a micro virtual machine secured by the company’s Titanium chip. That setup gives agents separate environments, rather than having them all share one runtime.
The service already powers Gemini. Google plans to open it for preview in the coming months, bringing the infrastructure to users beyond the system it currently supports.
Pausing idle agents frees capacity
An agent that is not doing work does not need to keep occupying an active virtual machine. Agent Compute Engine pauses an idle agent’s VM automatically, freeing the machine for other agents. When an agent needs to resume, it can return to its work environment.
Google also uses shared memory between VMs to reduce the cost of running each one. Virtual machines are a significant cost factor for agentic systems, so the combination of pausing and shared memory aims to make isolated agents more economical to operate.
A runtime for larger agent deployments
The same approach underpins Agent Substrate, an open-sourced Kubernetes runtime. Google says it can fit ten times as many agents on a single server. That capacity claim matters for deployments where many agents may be active at different times and would otherwise compete for machine resources.
In a test involving 200,000 agents, suspend and resume took less than 500 milliseconds, according to Google. The result speaks to how quickly the system can shift resources as agents become idle or active. It is a company-reported test, rather than a general performance guarantee for every deployment.
Controls for managing agents
Google is also adding tools to help operators oversee agent activity and access. These include token tracking, model routing, and support for third-party identity providers, including Ping Identity and SailPoint.
The controls are intended to extend beyond Agent Compute Engine. Google says they will also apply to Cloud Run, GKE, and managed endpoints, giving teams oversight features across several parts of its cloud environment.
Together, the infrastructure and management features address two practical needs of agentic systems: isolating agents from one another and coordinating resources as their workloads change. The preview will give users a chance to assess the service and its controls in their own deployments.