Google shows how GKE can start 300 agent sandboxes each second

Google showed changes to GKE Dataplane V2, the network layer of its Kubernetes software. Kubernetes controls groups of computers. The changes help a cluster operate very many short sandboxes, which are isolated areas where AI agents operate code. Google tells that GKE can start 300 isolated sandboxes each second, in less than one second, if network identity does not use metadata that changes.

How it works

GKE gives each pod, a small group of containers, a network identity from its labels. If each pod has different labels, such as timestamps, each pod must have its identity. A cluster can hold only 65,536 identities. Google tells teams to move these fields to annotations, which Dataplane V2 does not use, or to set Cilium rules that ignore them.

The numbers

Google tells that Agent Substrate on GKE, a tool that restores agent condition, held 500 activations each second with 200,000 agents. Stop and start again was less than 500 milliseconds. A customer that Google did not name operates hundreds of thousands of safe environments.

What is not known

All numbers come from Google. It gave no test information and did not name the customer. Google tells teams to keep in labels only the fields that are important for identity.

Insight: Google changes the limit in configuration and sells scale above it

On 10 November Google showed an open Kubernetes stack for AI, with a sandbox tool, gVisor,. At this time it gives a configuration change for the identity limit. For more scale it has Agent Substrate, which it tells bypasses some Kubernetes limits. This reads as a plan to repair the same bottleneck in the open tool, then keep the customer with its tool above it. Look at which layer you use.