NVIDIA details OpenShell agent guardrail architecture

NVIDIA describes the OpenShell Reference Architecture as a full-stack approach to agent development and operations at industrial scale.[1] OpenShell itself is an open-source runtime for building security and governance into agentic systems.[2] Its access model gives an agent what it needs, when it needs it.[2] NVIDIA says the architecture enables enterprises to build, contain, observe and enforce safe agents across infrastructure and edge systems.[1] For teams assessing agent deployments, the distinction matters: OpenShell is the runtime, while the reference architecture brings it into a broader infrastructure design.

NVIDIA says the architecture uses Vera Rubin systems for model inference, Vera CPUs for agent harnesses and tool use, and BlueField-4 DPUs running Sentry for enforcement.[1] Sentry monitors agents and enforces policies governing access to models, tools and data.[3] NVIDIA says Sentry runs on BlueField-4 DPUs, providing runtime protection isolated from the host, and enforces granular access controls in silicon.[3] The company says Sentry can contain agents within milliseconds.[3] These are vendor claims about design and performance, not independent evidence of outcomes in every deployment.

CoreWeave measured more than 3x faster sandbox startup, while NVIDIA says CoreWeave measured 1.8 times faster agentic-loop completion on Vera than on a leading x86 CPU.[2] NVIDIA also says a single rack can support more than 11,000 concurrent agent environments.[2] The company says Vera CPUs have 88 Olympus cores and deliver up to twice the performance of x86 on critical agentic tasks.[2] These figures can inform evaluation plans, but the evidence does not specify test conditions or show that the reported results will carry over to other workloads.

The OpenShell Reference Architecture is available now.[4] NVIDIA says it runs on enterprise operating systems including Canonical, Red Hat, Microsoft Windows and Microsoft Execution Containers.[1] The architecture also extends to RTX Spark, DGX Spark and DGX Station edge systems.[1] ServiceNow and SAP are building on the architecture, while BBVA and Santander are exploring it.[4] NVIDIA’s announcement lists Anthropic, AWS, Cisco, CrowdStrike, Dassault Systèmes, Dell Technologies, Google Cloud, HPE, IBM, Meta, Microsoft and OpenAI.[1] The architecture is part of NVIDIA’s broader Open Agent Safety platform, which NVIDIA launched after rogue-agent incidents involving companies including OpenAI.[5]