Google made Agent Sandbox on Google Kubernetes Engine generally available on May 20. The announcement addresses the infrastructure behind AI agents that execute code: how to start isolated environments quickly, pause them when idle and resume work without keeping every environment fully active.
Alongside the release, Google introduced Agent Substrate, a separate open-source project exploring how to support much larger numbers of short-lived agent tasks. The availability announcement applies to GKE Agent Sandbox; the Substrate project is an additional development effort.
Warm pools trade idle capacity for faster starts
Agent Sandbox maintains ready environments in a warm pool so each request does not need to wait for a fresh instance to start. Google says its integrated pool can allocate 300 sandboxes per second per cluster, with 90% of allocations completing within 200 milliseconds.
Those measurements describe sandbox allocation, not the time required for an agent to finish a task. The model, its tools and the work being executed add their own delays.
Pod snapshots provide the complementary mechanism. An idle workload can be suspended and resumed later, reducing the need to keep computing resources running while an agent waits for a person or event. Google also describes suspended virtual-machine capacity that can replenish the ready pool at lower cost.
For an application team, the appropriate pool size depends on its traffic pattern. Keeping more environments ready can reduce startup delays while increasing the resources reserved between requests.
Isolation remains part of the application design
Google says Agent Sandbox supports gVisor and default-deny Kubernetes network policy, with interfaces for other isolation systems such as Kata Containers. These controls concern where untrusted code runs and which network connections it can make.
They do not decide which business action an agent should be allowed to take. An application still needs to limit access to its data and external tools. Our explanation of AI agents and their tools separates the model’s suggested action from the application’s authority to perform it.
Agent Substrate aims to reduce scheduling overhead for very large numbers of brief tasks while retaining Kubernetes infrastructure underneath. Google invites community work on that project, rather than presenting it as a completed replacement for the generally available sandbox product.
