An AI agent that runs code may work for a few seconds, wait for a model or a person, then work again. Keeping a dedicated machine or container busy through every pause can waste resources. Google Cloud announced Agent Substrate 1.0 and general availability on GKE on November 5, 2026 as an infrastructure answer to that pattern. The company also said Agent Substrate, gVisor and OpenRL had been accepted into the CNCF Sandbox and were beginning onboarding. Acceptance is a governance step, not proof that every claimed performance figure has been independently reproduced.
How does Agent Substrate keep an agent's work between turns?
The runtime separates an Actor, which holds an agent's identity, memory and files, from a Worker, the sandbox that executes it. A warm Worker pool serves active requests. When an Actor waits, the system snapshots its state to storage and frees that Worker; a later request restores the Actor onto an available Worker. A Gateway controls access and injects secrets according to Actor identity. Google says this can resume a suspended Actor in under 500 milliseconds and reports high sandbox density at scale. Those timings and density figures are Google-reported, not a promise for a new cluster.
The isolation boundary can use gVisor or a Cloud Hypervisor microVM. This is significant for coding agents because they may execute generated commands or packages that have not been reviewed. The architecture gives an app builder a way to keep each agent's files and permissions separate while recycling compute during idle periods. It does not remove the need to decide which tools an agent may call or how to review consequential actions.
Can any Kubernetes or GKE cluster use it today?
Google describes two routes: managed Agent Substrate on GKE and an open-source deployment on a Kubernetes cluster. Its GKE documentation currently requires a Standard cluster on version 1.37 or later with Workload Identity Federation. Autopilot is not supported. The microVM route needs suitable x86 nodes with nested virtualization or bare metal, and the documentation says GPU device passthrough into Actor containers is not supported. An agent that maintains database or tool connections must reconnect after suspension because open network connections are not preserved.
What happened to gVisor and OpenRL?
The November 5 announcement places three separate projects in one infrastructure story. Agent Substrate manages agent lifecycles and sandboxes; gVisor supplies a user-space kernel isolation option; OpenRL is a Python-oriented system for reinforcement-learning runs on Kubernetes. Google says all three have entered the CNCF Sandbox process. They are not a single product a maker must adopt together, and a small app does not need a million-agent architecture to benefit from safe execution boundaries.
The immediate reader question is whether a stateful agent app needs suspension, per-agent isolation and a reusable execution pool. If so, the new 1.0 release and its published constraints give a concrete option to evaluate. VIBPS is the social network for software and has no role in the project; this article is based on Google's announcement and documentation, not a deployment test.