NVIDIA Corporation (Nasdaq: NVDA) has introduced its OpenShell Reference Architecture, a design for running enterprise AI agents with runtime isolation, continuous monitoring and policy enforcement in the underlying infrastructure. The company said the system is available now and is intended to give organizations a way to deploy agents while retaining technical controls over what those agents can access and do.
The architecture extends NVIDIA OpenShell secure-runtime software into a stack for building and operating agents. It combines Vera Rubin systems for model inference, Vera CPUs for agent harnesses and tool use, and BlueField-4 data processing units running NVIDIA Sentry. NVIDIA said Sentry monitors agent behavior outside the main compute path and can quarantine and stop an agent in milliseconds if it attempts to move beyond its assigned software boundary.
That boundary addresses a practical deployment problem for enterprise IT teams. An agent that can call tools, handle data or take actions across systems needs controls that continue to operate after the model generates a response. NVIDIA is positioning the reference architecture as an infrastructure layer that can contain agents, observe their behavior and enforce policies independently of the agent application itself.
The company said BlueField-4 and Sentry provide in-silicon policy enforcement, while the secure runtime supplies a controlled operating boundary for autonomous agents. NVIDIA also said the design can support both closed and open models, and that its Vera Rubin NVL72 infrastructure can be used across pretraining, post-training and inference workloads. The company reported that Sentry can increase agent capacity by as much as 20%; that figure is an NVIDIA performance claim and was not independently verified.
For platform teams, the architecture reaches beyond a single hardware configuration. NVIDIA said it runs with Linux and Kubernetes infrastructure as well as a Windows container environment, and can operate on public-cloud systems. That compatibility matters because agent workloads are likely to span data centers, cloud services and edge deployments rather than remain in one standardized environment.
NVIDIA said the design also extends to edge systems using RTX Spark, DGX Spark and DGX Station products. It identified potential uses in retail, healthcare, manufacturing, robotics, software development and telecommunications. The company named a broad set of infrastructure, cybersecurity and enterprise-software organizations that are working with the architecture, but the source does not establish the scope or production status of each integration.
The immediate operational benefit is a reference design that places agent containment and monitoring alongside the compute and networking components used to run the workload. That could simplify the task of defining technical boundaries for agents that access tools or sensitive systems, though organizations would still need their own identity policies, access controls and application-level governance.
As enterprises move from AI experiments to longer-running agents, the operational constraint is no longer only model performance. Security teams need controls that persist at runtime and can interrupt unsafe activity. NVIDIA’s architecture offers one approach centered on infrastructure enforcement, while its practical value will depend on how well enterprises integrate it with existing security and operations processes.