At KubeCon + CloudNativeCon, NVIDIA announced that its OpenShell runtime and the Google‑originated Agent Substrate framework are now CNCF sandbox projects, and it opened public preview of three DSX AI Factory components—NVKE, NVKR and Snapshot. For anyone operating AI workloads on Kubernetes, the move adds community‑governed code for agent security, scalable agent execution, and rapid model warm‑up.
OpenShell becomes a CNCF sandbox project
OpenShell is an open‑source runtime that enforces deterministic policies on autonomous agents. It lets teams define what an agent may do, how it accesses inference services, and which credentials it can use. The runtime’s plug‑in architecture can be tied into existing compute back‑ends, inference providers, secret stores and monitoring pipelines. According to the announcement, hundreds of organizations have already adopted OpenShell and the project has gathered tens of thousands of contributions within its first year.
Agent Substrate for large‑scale agent workloads
Agent Substrate, now also in the CNCF sandbox, provides a framework for running long‑lived agent workloads on Kubernetes. It integrates with NVIDIA’s Dynamo inference layer and the KAI Scheduler for shared‑cluster resource allocation. The design preserves an agent’s environment between tasks and releases compute when the agent is idle, allowing a pool of agents to be suspended and resumed on demand. NVIDIA and Google are collaborating on a joint integration that would let OpenShell policies be applied to Agent Substrate workloads across cloud and on‑prem environments.
New DSX AI Factory open‑source components
The DSX AI Factory Platform now includes three publicly previewed projects. NVKE (NVIDIA Kubernetes Engine) and NVKR (NVIDIA Kubernetes Recipes) automate the lifecycle of GPU‑enabled clusters and ship signed, version‑pinned images. A list of ecosystem partners—including Mirantis, Netris, OpenNebula Systems, Rafay, Spectro Cloud, vCluster, Firmus and SK Telecom—has committed to integrating these components into their offerings. Snapshot addresses the latency of loading models and warming up workloads. It restores checkpointed replicas in seconds, avoiding a full model reload. Design partners Crusoe and Saturn Cloud reported reductions in start‑up time from several minutes to under a minute when using Snapshot on NVIDIA HGX B300 hardware.
Related CloudNinjas coverage: AI engineering.
What This Means For Practitioners
Practitioners should evaluate whether their agent‑centric pipelines can benefit from OpenShell’s policy engine and consider adopting Agent Substrate for workloads that need to be paused and resumed at scale. Teams deploying GPU clusters on Kubernetes can trial NVKE and NVKR to standardize cluster provisioning and image management, especially if they already work with any of the listed partners. Finally, incorporating Snapshot may reduce cold‑start latency for inference services, which can improve elasticity and recovery times in production. Monitoring the CNCF sandbox progress and the public preview status of these projects will be essential for planning integration timelines.
