1. News
  2. Cloud & infrastructure
  3. Google Cloud · AI agents
Google Cloud · AI agents2 min read

Google unveils a safer, separate computer for each AI agent it runs

Google says each AI agent it runs for Gemini gets its own walled-off computer, and the tools to control agent access and cost are expanding.

By the Dashline deskSource: Google Cloud blog
Google unveils a safer, separate computer for each AI agent it runs
Photo: Victorgrigas, CC BY-SA 3.0

Google says every AI agent it runs for Gemini now gets its own separate, locked-down computer. The service is called Agent Compute Engine, and it was announced at the AGNTCon + MCPCon event. It already powers Gemini as an agent for work tasks. A preview for outside customers is due in the coming months.

What changed

An AI agent is software that carries out tasks on its own, such as writing and running code, using tools or browsing the web. Google says ordinary computing setups were not built for that fast, repeated activity. Agent Compute Engine gives each agent its own sandbox, a walled-off space, with isolation built into the hardware. Google says businesses will soon be able to use it to run their own custom agents.

A version for Kubernetes users

Google also described Agent Substrate, an open-source tool for companies that already use Kubernetes, a common system for managing software across many machines. When an agent goes idle, the tool saves its active memory and frees the machine for other work. It then resumes the agent when needed. Nous Research says it uses this approach to deploy agents inside customer systems.

Keeping agents in check

Google names the risks of unmanaged agents: shadow AI, which means unapproved AI use, data leaks, and actions that are hard to track and audit. It is extending the agent governance tools in Gemini for business to agents on Google Cloud, other clouds, company-owned servers and client devices. L'Oréal says this gives it clear control over what each agent can access. New Agent Gateway features arrive in preview this quarter:

  • Granular cost controls
  • Intelligent model routing
  • Simpler customizations and integrations

What is not known yet

The announcement as published here gives no prices and no exact preview dates for Agent Compute Engine. It also does not spell out the full feature lists. For any business planning to use agents, the questions to ask are who controls what they can reach, what they cost, and how their actions get recorded.

The takeaway

Before adopting AI agents, ask any vendor how each agent is isolated, what it can access, what it costs, and how its actions are logged.

More articles

You should really read these.

Next step

Let’s talk