GKE Agent Sandbox GA and Agent Substrate Launch

AI agents are evolving beyond simple chat interfaces to autonomously execute code and use tools. This transition introduces a critical infrastructure challenge: securely running untrusted code at scale with low latency. Google Kubernetes Engine (GKE) Agent Sandbox addresses this by providing a cloud-native, open-source execution environment designed specifically for agent workloads. The solution has seen rapid community adoption, with over 16x growth in sandbox deployments in five months and integrations with partners like LangChain and Lovable.

Technically, GKE Agent Sandbox tackles performance bottlenecks inherent to traditional Kubernetes setups for agents. It integrates Pod Snapshots to suspend idle agents, drastically reducing compute waste. The new Sandbox API with an integrated warm pool enables allocating 300 sandboxes per second per cluster, with 90% of allocations completing in under 200 milliseconds. A cold pool of suspended VMs further reduces warm pool maintenance costs. Security is ensured through native support for gVisor and default-deny network policies, with pluggable isolation options like Kata Containers. Running on Axion processors, it offers up to 30% better price-performance than comparable hyperscalers.

While Agent Sandbox is now generally available and production-ready, Google is also introducing Agent Substrate, a new open-source project. Agent Substrate aims to push infrastructure density and performance limits for ultra-scale agentic workloads, which can reach hundreds of millions of instances. It pairs the secure runtime and snapshotting of Agent Sandbox with a minimal control plane that bypasses standard Kubernetes control plane bottlenecks for high-frequency, sub-second tool calls. This new abstraction layer is designed for the future of autonomous agents, inviting community collaboration to build the next chapter in agent-native infrastructure.