Google gives its Agent Substrate sandbox for AI agents to a foundation

Google Cloud released version 1.0 of Agent Substrate and gave it to the Cloud Native Computing Foundation. It puts AI agents, which are software that does tasks, in sandboxes on Kubernetes, software that controls groups of computers. Google engineers spoke about it at KubeCon + CloudNativeCon North America. They tell that agents must have a different cluster set-up from other software.

How it works

Agents write code. Google tells that this code is not safe until a check is done. A sandbox isolates it, and network guardrails limit where it can go. Google also tunes each cluster for the work that it operates.

The proof

Google gave no test results. One customer problem is clear. Agents are not active for most of the time, but a sandbox stays in use for the full session. Jaana Dogan, an engineer at Google Cloud, tells that many processor cycles are lost.

What is not known

Google gave no number for the savings from Agent Substrate. Jago Macleod, a director at Google Cloud, tells that agent work is too much work for the Kubernetes control plane. That is the part that controls the cluster. Some teams made other control planes.

Insight: Google fixes Kubernetes in layers for agents

Google also showed six open AI tools for Kubernetes, an update for the Python client, and a network change for 300 sandboxes each second. This reads as a race to keep Kubernetes as the platform for agents, and not a new stack that Google does not shape. Google tells that customers did not want agents to divide their set-up. Look for the same moves from other cloud providers.