News
NVIDIA outlined OpenShell, Vera CPUs and BlueField-4 as a hardware-rooted stack for governing AI agent workloads on Oct. 21, 2026.
NVIDIA on Oct. 21, 2026 detailed OpenShell Reference Architecture, a hardware-rooted stack designed to govern AI agent workloads that can operate for hours, access tools and the internet, and spawn other agents. The company said the architecture is intended to keep agent execution under infrastructure-level controls rather than relying on the agent itself.
Evidence and context
In the technical blog published Oct. 21, 2026, NVIDIA said autonomous agents create compounding risk as they reason through tasks, download code, request privileges and touch more of the underlying environment over time. The company framed OpenShell as an open-source runtime that defines and enforces which systems and data an agent can access, including files, internet services and systems connected to physical devices.
NVIDIA said OpenShell treats agent workloads as untrusted throughout execution and is built around four components: Gateway, which manages sandbox lifecycles and policies; Supervisor, which checks outbound requests and mediates access to credentials; Sandbox, which isolates the workload and restricts filesystem and process access; and Policy Prover, which formally checks whether permissions remain within an organization-defined ceiling. The architecture also includes supervisor middleware and gateway interceptors for deployment-specific governance.
The company said the reference architecture runs on NVIDIA Vera CPUs, which it described as the compute layer beneath OpenShell. NVIDIA said Vera is optimized for agent workloads that depend on repeated loops of reasoning, tool calls, code execution and data queries. The blog cited 88 NVIDIA-designed Olympus cores, a second-generation Scalable Coherency Fabric, LPDDR5X memory, 96 lanes of PCIe Gen 6, confidential computing and an energy-efficient design for power-limited data centers.
NVIDIA attributed several performance claims to Vera. It said the platform delivers up to 2x performance on critical agentic workloads versus x86, up to 3x core-to-core bandwidth and up to 50% lower core-to-core latency than x86 alternatives, up to 1.2 TB/s of LPDDR5X memory bandwidth, up to 40% lower loaded memory latency than chiplet-based designs, and 2x more performance per watt than x86 CPUs. Those figures were presented as vendor measurements in NVIDIA’s blog, not independent benchmarks.
The company also cited CoreWeave testing of agent sandboxes on Vera, saying sandbox startup was more than 3x faster than on x86 and that a single rack supported more than 11,000 concurrent agent environments. NVIDIA said CoreWeave found Vera delivered 1.8x faster performance on Terminal Bench, a CPU benchmark for real agentic loop tasks spanning scripting to database recovery.
For additional protection, NVIDIA said OpenShell can be paired with BlueField-4 DPUs and NVIDIA Sentry, a security software framework built on DOCA. In the architecture, BlueField-4 sits in the I/O path between agents on Vera and the models they access, while Sentry monitors requests and responses, networking and file activity, verifies identity and delegated authority, and analyzes selected host memory. NVIDIA said BlueField-4 can enforce policy in silicon and quarantine an agent within milliseconds when a threat or policy violation is detected.
The company positioned OpenShell, Vera Rubin and Sentry on BlueField-4 as the pillars of the OpenShell Reference Architecture, and said the stack is meant to give organizations both performance and control for safe AI agents. The blog did not provide an external verification of the security guarantees or the measured timing claims.
