GKE Agent Sandbox is now generally available and Agent Substrate has been releases as an open-source project, announced

GKE Agent Sandbox is now generally available with stable APIs and features for secure, high-throughput agent workloads, and Agent Substrate has been open-sourced to push scheduling density and performance for large agent fleets.

GKE Agent Sandbox is now generally available and Agent Substrate has been released as an open-source project, announced May 21, 2026 by Brandon Royal and Tim Hockin. The GA move makes the sandbox a supported product with stable APIs for running agent workloads on Kubernetes, while Agent Substrate aims to increase the density and performance of agentic infrastructure. Together they form a foundation for deploying agents at much larger scale with predictable behavior and integration paths.

The GA release highlights features built for agent patterns. Pod Snapshots let clusters suspend idle agents and resume them in seconds. A Sandbox API with an integrated warm pool supports high-throughput allocations — reported at 300 sandboxes per second per cluster, with 90% of allocations completing within 200 milliseconds. Cost-saving standby capacity buffers use suspended VMs to replenish warm pools more cheaply than constant hot capacity.

Adoption accelerated after the preview at KubeCon NA in November 2025: the company reports more than 16× growth in sandboxes on GKE in under five months, and partnerships with customers such as LangChain and Lovable deploying millions of agents into production. Moving from preview to GA and stabilizing APIs is intended to position the sandbox as an infrastructure layer the broader agent ecosystem can build on.

For builders, GA promises lower idle compute and much faster sandbox start — up to handle the bursty, short — lived nature of many agents while retaining strong isolation. The platform is positioned to help teams manage fleets that scale toward tens or hundreds of millions of instances, reducing cold-start latency and preserving network and kernel isolation even under heavy scheduling pressure.

Agent Substrate is presented as the next step to push agentic infrastructure density further. The open-source project targets dense scheduling for workloads that are highly concurrent but often idle, enabling more aggressive suspend/resume behavior, tighter kernel and network isolation, and innovations designed to relieve pressure on the Kubernetes control plane. It is intended to complement the sandbox by enabling higher packing density and control — plane efficiency.

Security and customization are emphasized for production deployments. Agent Sandbox natively supports gVisor and default — deny Kubernetes network policy, and it exposes pluggable interfaces for alternative sandbox technologies such as Kata Containers. The announcement also cites up to 30% better price — performance on Axion processors versus comparable hyperscaler clouds, highlighting both performance and cost considerations for prospective deployments.