NVIDIA has announced an OpenShell Reference Architecture designed to help enterprises build, contain, observe and enforce safer artificial-intelligence agents. The company describes it as a full-stack approach rather than a standalone model feature. The design is intended for production environments where agents may handle sensitive data and invoke business tools. It combines computing, runtime software and security enforcement so organizations can give agents useful access to tools while keeping their actions inside defined boundaries. NVIDIA made the announcement at GTC Berlin on October 21.
The architecture brings together NVIDIA Vera Rubin infrastructure for model inference, Vera CPUs for agent harnesses and tool use, and BlueField-4 data-processing units running NVIDIA Sentry. OpenShell supplies the secure runtime boundary, while Sentry continuously monitors behavior and applies policy enforcement in silicon. NVIDIA says an agent that moves outside its software boundary can be quarantined and stopped in milliseconds, making control part of the infrastructure layer. The goal is to make containment measurable and persistent instead of relying on a one-time configuration or a model prompt.
NVIDIA says the design can extend from enterprise data centers to edge systems using RTX Spark, DGX Spark and DGX Station platforms. It supports leading operating systems and cloud environments, including Linux and Kubernetes infrastructure, Microsoft Windows with Microsoft Execution Containers, Amazon Web Services, Google Cloud and Microsoft Azure. The company positions the architecture for sectors such as healthcare, retail, manufacturing, robotics, software development and telecommunications. That breadth shows NVIDIA is targeting operational systems, not only research labs. The architecture is therefore aimed at practical control as companies move from pilots to always-on services.
The announcement also lists model and technology partners working with OpenShell, including Anthropic, Google, Meta, OpenAI, Perplexity, Microsoft, Cisco, CrowdStrike, SAP and ServiceNow. That partner list signals a push to make agent governance compatible with multiple providers rather than locking safety to one model. Enterprises will still need their own access rules, approvals and monitoring, but NVIDIA’s pitch is that safeguards should be enforced continuously across software and hardware. Enterprises will still need their own access rules, approvals and monitoring, but NVIDIA’s pitch is that those safeguards should be enforced continuously across software and hardware.
