NVIDIA announces the OpenShell reference architecture for agents
A full-stack design for running agents on Vera Rubin, Vera CPUs, and BlueField, with Sentry to stop an agent that leaves its boundary. NVIDIA says the architecture is available now.
Official source (opens in a new tab)
NVIDIA's newsroom post on 21 October 2026, from GTC Berlin, announces the OpenShell Reference Architecture. It is a design for building, containing, watching, and enforcing rules on agents across enterprise machines and edge systems. It extends the existing OpenShell secure runtime. NVIDIA's stack pairs Vera Rubin systems for inference, including Nemotron and closed models, with Vera CPUs for the agent harness and open-source OpenShell software. BlueField-4 DPUs run NVIDIA Sentry, which NVIDIA says watches agent behavior out of band and enforces policy in silicon. If an agent leaves its software boundary, NVIDIA says Sentry quarantines it and stops it in milliseconds.
NVIDIA says Vera Rubin NVL72 delivers 30 times the throughput per megawatt on agent workloads, that Vera CPUs deliver twice the performance of x86 processors, and that Sentry can raise agent capacity by up to 20 percent. The design also covers RTX Spark, DGX Spark, and DGX Station at the edge, and NVIDIA says it runs on Canonical, SUSE, Red Hat, and Microsoft Windows with Microsoft Execution Containers, plus the cloud operating systems from Amazon Web Services, Google Cloud, and Microsoft Azure. NVIDIA says enterprises can run agents from builders including Anthropic, Google, Meta, OpenAI, and Perplexity on this infrastructure. NVIDIA says the reference architecture is available now.
Who it's for: an infrastructure team that has to run many agents and wants the containment story in the hardware, not only in the application.
Fits: Best Agent Builder stays n8n, the product a team uses to design a workflow. OpenShell is the machine those agents can run on. The verdict is unchanged.