NVIDIA launches the OpenShell Reference Architecture for safe AI agents at GTC Berlin

NVIDIA announced the OpenShell Reference Architecture at GTC Berlin on 21 October 2026 alongside version 0.1.x of the open-source NVIDIA OpenShell runtime under the Apache License 2.0. Treating agent workloads as untrusted throughout execution, OpenShell separates governance into four components outside the agent’s reach: a Gateway control plane, a Sandbox that restricts filesystem and system-call access, an out-of-sandbox Supervisor that checks outbound network traffic and injects credentials only for approved endpoints, and a Policy prover that uses formal verification to check whether a policy change stays within an organisation’s ceiling.

The full-stack architecture pairs OpenShell with NVIDIA Vera Rubin NVL72 infrastructure for model inference, NVIDIA Vera CPUs (with 88 Olympus cores and up to 1.2 TB/s of LPDDR5X memory bandwidth) for running agent harnesses, and NVIDIA BlueField-4 DPUs running NVIDIA Sentry on DOCA to monitor behaviour out of band and quarantine out-of-bounds agents in milliseconds. The stack also extends to the edge on RTX Spark, DGX Spark and DGX Station systems.

In benchmarks run by CoreWeave on NVIDIA Vera CPUs against x86 servers, sandbox start-up was more than three times faster, supporting more than 11,000 concurrent agent environments in a single rack, and Terminal Bench agentic loops ran 1.8 times faster.

Companies working with the architecture include Anthropic, Amazon Web Services, Cisco, CrowdStrike, Google Cloud, Meta, Microsoft, OpenAI, Palo Alto Networks, Perplexity, Red Hat, SAP, ServiceNow and SpaceX Super Intelligence.