Nvidia donates OpenShell to CNCF as it backs agent scale

NVIDIA has contributed its OpenShell project to the Cloud Native Computing Foundation, placing software for governing autonomous AI agents into the CNCF Sandbox.

It is also contributing to Agent Substrate, a Google-led framework for running long-lived agent workloads on Kubernetes, and has introduced three open source additions to its DSX AI Factory Platform: NVIDIA Kubernetes Engine, NVIDIA Kubernetes Recipes and NVIDIA Snapshot.

Kubernetes has become a central layer for managing AI workloads across large computing environments, and NVIDIA is seeking a bigger role in the open source software stack around it. The latest changes focus on two issues that have grown more pressing as AI systems spread: how to govern autonomous agents and how to manage the infrastructure needed to run them at scale.

Agent governance

OpenShell is designed to control what AI agents are allowed to do as they use tools, access inference services and execute workloads. NVIDIA says the software applies deterministic controls to agent activity and can connect with an organisation's own compute, inference providers, credential storage and monitoring tools.

By moving the project to the CNCF Sandbox, NVIDIA is placing its governance under a broader community within the Linux Foundation. The approach is intended to let product makers and security specialists inspect and improve the software openly.

According to NVIDIA, hundreds of organisations are already using OpenShell, and the project has attracted tens of thousands of contributions in less than a year since launch.

Jim Zemlin, Chief Executive Officer of the Linux Foundation, described the contribution as a response to a growing security issue in cloud-native software.

"The cloud-native community has a new infrastructure challenge: giving agents secure environments in which to operate," said Jim Zemlin, Chief Executive Officer of the Linux Foundation.

"NVIDIA's contribution of OpenShell to CNCF gives developers a shared foundation to build on and an opportunity to shape how agent security evolves. Bringing this work into the community enables broader collaboration on infrastructure that will be critical to the next generation of cloud-native applications," Zemlin said.

Scaling agents

NVIDIA's work on Agent Substrate addresses a separate operational issue: how to run large numbers of agents without dedicating computing resources to every session at all times. The framework preserves agent environments between tasks while releasing compute resources when those agents are idle.

This suspend-and-resume model is intended to help users scale shared pools of agents more efficiently on Kubernetes. NVIDIA is working to ensure the framework operates with NVIDIA Dynamo for model inference and KAI Scheduler for shared cluster resource allocation.

NVIDIA is also collaborating with Google to integrate OpenShell and Agent Substrate so businesses can apply agent governance across both cloud and on-premises systems.

Infrastructure layer

Alongside its agent-related work, NVIDIA unveiled new open source infrastructure components tied to its DSX AI Factory Platform. They are aimed at organisations that want managed Kubernetes services on GPU-based systems while retaining control of their own infrastructure and software environment.

NVIDIA Kubernetes Engine, or NVKE, and NVIDIA Kubernetes Recipes, or NVKR, are available in public preview. The projects automate cluster lifecycle management and provide signed, version-pinned images for Kubernetes deployments on GPU infrastructure, according to NVIDIA.

Several companies have aligned with NVIDIA on planned integrations for the tools, including Mirantis, Netris, OpenNebula Systems, Rafay, Spectro Cloud, vCluster, Firmus and SK Telecom.

NVIDIA Snapshot, another new open source component, targets a common bottleneck in AI inference systems: the time needed to load models and warm up workloads before extra capacity can be brought online. According to NVIDIA, the software restores checkpointed replicas in seconds without requiring model reload or warmup.

NVIDIA cited early testing by design partners Crusoe and Saturn Cloud. Crusoe, which contributes to Snapshot, tested the software with GLM-5.3 on Crusoe Cloud and reduced workload readiness time from about eight minutes to 31 seconds, according to NVIDIA. Saturn Cloud said it cut vLLM startup times for scale-from-zero inference from between two and 8.5 minutes to six to 34 seconds.

The announcements show NVIDIA extending its influence beyond chips and model software into the operational systems used to run AI services. With Kubernetes now widely used as the orchestration layer for cloud-native applications, the company is betting that open source software for agent control, inference recovery and GPU cluster management will become a more important part of the AI infrastructure market.