Nvidia has launched the OpenShell Reference Architecture, a system for running AI agents with controls over what they can do. The company says it combines its OpenShell secure runtime with computing infrastructure that can monitor agents and enforce security policies.
OpenShell provides a software boundary for an agent. Nvidia says its BlueField-4 DPUs run Sentry security software outside that boundary to monitor behavior. If an agent tries to cross it, Sentry can quarantine and stop the agent in milliseconds, according to the company.
The architecture combines chips and software
The architecture uses Vera Rubin infrastructure for model workloads and Vera CPUs for agent harnesses and tool use. Nvidia says Vera Rubin NVL72 systems deliver 30 times higher throughput per megawatt for agent workloads, while Vera CPUs deliver twice the performance of x86 processors. It also says Sentry running on BlueField-4 can increase agent capacity by up to 20%.
Nvidia says the architecture runs on Linux and Kubernetes infrastructure powered by Canonical, SUSE and Red Hat, as well as Microsoft Windows using Microsoft Execution Containers. It also extends to edge systems including RTX Spark, DGX Spark and DGX Station.
Read nextMiniMax launches M3.1 Flash, claims 70% lower AI agent costsPartners are working on deployments
Nvidia named cloud providers, security companies, model builders and enterprise software makers working with the architecture. SAP is integrating it with its Joule Studio runtime. Nvidia says BBVA and Santander are exploring it as a framework for agent security and governance.
Nvidia cautions that many products and features it describes are still in development and may be offered only if they become available. The OpenShell Reference Architecture itself is available now.
