Details
- NVIDIA announced the OpenShell Reference Architecture on October 21, 2026, providing infrastructure for building, containing, monitoring and enforcing enterprise AI agents.
- The platform combines Vera Rubin systems for model inference, Vera CPUs for agent runtimes and tool use, OpenShell secure runtime software, and BlueField-4 DPUs running NVIDIA Sentry for hardware-level policy enforcement.
- OpenShell isolates agents while Sentry independently monitors requests, responses, network activity, file access and delegated authority; policy violations can trigger quarantine within milliseconds.
- The architecture supports enterprise and edge deployments across Linux, Kubernetes and Windows environments, with systems including NVIDIA RTX Spark, DGX Spark and DGX Station extending coverage beyond data centers.
- NVIDIA lists partnerships spanning model providers, cloud companies, cybersecurity vendors, systems manufacturers and enterprise software firms; ServiceNow also made its Arc long-running agent generally available with runtime enforcement through OpenShell.
Impact
The announcement reflects a shift toward treating agent security as an infrastructure problem rather than solely a model or application concern. If adopted, independent monitoring and silicon-enforced controls could support longer-running enterprise agents while influencing procurement, cloud architecture and governance standards over the next 12–24 months.