Nvidia outlined a wider set of open source tools for AI infrastructure on November 10, combining its contribution of the OpenShell agent runtime to the Cloud Native Computing Foundation with new Kubernetes software for GPU operators. The announcements address how companies control agents and keep the computing systems behind them running.
Kubernetes is software for coordinating applications across groups of computers. An agent runtime is the environment in which an AI system uses tools and carries out tasks. The two serve different purposes, but both become important when a prototype agent moves into a shared business environment.
Controls outside the model
CNCF welcomed OpenShell at KubeCon + CloudNativeCon North America, describing a system that enforces boundaries on agents from outside the agents themselves. The foundation said OpenShell is adopting Cedar as its policy language and identified contributors from Nvidia, Red Hat, Docker and Canonical.
The distinction matters because a written instruction to an AI model is different from a technical restriction on what its software can access. A company may want an agent to summarize a document while preventing it from reaching unrelated files or network services. That restriction needs an enforceable boundary, not merely an expectation of good behavior.
CNCF’s project record lists September 26, 2026 as OpenShell’s acceptance date at the Sandbox maturity level. The November 10 event therefore marks its public welcome and broader infrastructure announcement, rather than evidence that the project was first accepted that day. Sandbox is an early project stage, not a certification that every deployment is safe.
New tools for GPU infrastructure
Nvidia also announced public previews of Kubernetes Engine and Kubernetes Recipes, plus Snapshot, a component intended to restore saved inference workloads without reloading a model. Nvidia says it is contributing to Google’s Agent Substrate project and working on integration between that system and OpenShell.
The Kubernetes Engine repository describes a management system for operators running GPU infrastructure on their own machines. Customers receive separate workload clusters, while the operator runs the management layer. Its companion Recipes project produces the node images and components that the engine deploys.
Preview limits require attention
The repository explicitly advises against production workloads during preview because APIs, configuration and behavior can change. It also says authentication is initially off and the API uses plain HTTP until the operator configures protected access. Those are material setup requirements for anyone evaluating the software.
Other documented limits include the absence of a management-plane upgrade or backup-and-restore procedure. Nvidia Infra Controller is the only compatible infrastructure provider in the preview. These constraints make the release an evaluation opportunity, not a drop-in replacement for an established production platform.
For enterprise teams, the next step is to assess the components separately: agent permission enforcement, cluster management and workload recovery solve different problems. Wider community participation can improve the tools, but organizations still need to test their own policies and operating procedures before connecting sensitive systems.
