NVIDIA announced the OpenShell Reference Architecture at GTC Berlin on October 21, 2026, a design for building, running and monitoring AI agents on NVIDIA hardware, and says it is now available. It builds on NVIDIA's open source OpenShell runtime. Its hardware monitoring layer, NVIDIA Sentry, is not out yet: NVIDIA's developer blog describes "NVIDIA Sentry’s planned availability early next year."

#What changed?

According to NVIDIA's announcement, the architecture brings together "NVIDIA Vera Rubin infrastructure for model inference, NVIDIA Vera CPUs for agent harnesses and tool use, and NVIDIA BlueField-4 DPUs running NVIDIA Sentry for in-silicon enforcement." NVIDIA says it extends OpenShell secure runtime software into infrastructure for agent development and operations, and that it is extensible to agents at the edge on NVIDIA RTX Spark, DGX Spark and DGX Station systems.

"As enterprises prepare to deploy billions of AI agents, safety and control must be built into the computing architecture," said Jensen Huang, founder and CEO of NVIDIA, in the release.

SDxCentral reported that Ian Buck, whom it identifies as NVIDIA's VP and GM of hyperscale and high-performance computing, said in a press briefing: "Long-running agents combine tools, credentials, and network access in ways that traditional security controls were not designed to govern."

#How does OpenShell work?

"NVIDIA OpenShell is an open-source runtime to build the security and governance needed for agentic systems into the computing infrastructure," NVIDIA says in a developer blog post. "OpenShell treats agent workloads as untrusted throughout execution." NVIDIA lists four components:

Component

What NVIDIA says it does

Gateway

"The control plane that manages sandbox lifecycles and policies"

Supervisor

"Checks outbound requests against policy and mediates access to credentials"

Sandbox

"Isolates the agent workload and restricts filesystem and process access"

Policy prover

Checks "whether a policy’s permissions stay within an organization-defined ceiling"

"OpenShell separates policy enforcement from the agent workload, so a compromised agent can’t change its own restrictions," NVIDIA says. In the reference architecture, OpenShell runs on NVIDIA Vera CPUs.

#What will NVIDIA Sentry add?

NVIDIA describes Sentry as "a security software framework that monitors AI agents and enforces policies governing their access to models, tools, and data," built on NVIDIA DOCA and running on BlueField-4 DPUs. In a separate developer blog post, NVIDIA says BlueField-4 "sits on the node’s only path to the model," and that Sentry enforces access policies "without relying on the agent to report its own behavior."

NVIDIA says Sentry can "Govern agent-to-agent (A2A) and Model Context Protocol (MCP) traffic" and "Detect and redact personally identifiable information and secrets, including API keys, tokens, and passwords." Its parts include DOCA Tower, an AI security gateway built on Envoy; a per-agent layer-4 firewall built with DOCA Flow; the Open Policy Agent policy engine; and DOCA Argus for live memory introspection. NVIDIA also claims that "Even if an agent escapes its sandbox, NVIDIA Sentry retains its monitoring and enforcement capabilities."

NVIDIA calls Sentry on BlueField-4 an "optional hardware safeguard alongside OpenShell’s software boundaries."

#What performance does NVIDIA claim?

All of these are NVIDIA's claims, with no baselines given for the first and third:

  • "For agentic workloads, NVIDIA Vera Rubin NVL72 systems deliver 30x higher throughput per megawatt."
  • Vera CPUs deliver "up to 2x performance on critical agentic workloads versus x86," according to NVIDIA's developer blog.
  • Sentry "boosts agent capacity by up to 20%."
  • NVIDIA says CoreWeave found "sandbox startup to be more than 3x faster on Vera than on x86, with a single rack supporting more than 11,000 concurrent agent environments."

#Who can use it and when?

Part

Status, per NVIDIA

OpenShell Reference Architecture

"now available"

NVIDIA Sentry

"planned availability early next year" (developer blog)

DOCA Argus, DOCA Flow, OVS-DOCA

Available "for experimenting" ahead of Sentry

Price

Not published

NVIDIA says the architecture runs on Linux and Kubernetes infrastructure from Canonical, SUSE and Red Hat, on Microsoft Windows using Microsoft Execution Containers, and on cloud operating systems from Amazon Web Services, Google Cloud and Microsoft Azure.

NVIDIA's release lists partners by role:

  • Model builders: enterprises can run agents from Anthropic, Cohere, Google, Meta, OpenAI, Perplexity, Reflection and SpaceX Super Intelligence on OpenShell infrastructure.
  • Security vendors: 1Password, Check Point, CrowdStrike, Cisco, F5, Fortinet, Microsoft, Okta, Palo Alto Networks, TrendAI and Zscaler are integrating support.
  • System makers: Dell Technologies, Cisco, HPE, Lenovo and Supermicro are among those building infrastructure with it.
  • Banks: BBVA and Santander "are working with NVIDIA to explore the OpenShell Reference Architecture as a framework for strengthening agent safety, security and governance."
  • Software: SAP is integrating it with Joule Studio runtime, and NVIDIA says ServiceNow announced general availability of ServiceNow Arc, an agent "enforced by NVIDIA OpenShell."

#ServiceNow Arc goes GA on OpenShell

ServiceNow announced general availability of Arc on October 21, 2026, at NVIDIA GTC in Berlin. NVIDIA's release describes Arc as "a personal long-running AI agent with actions governed at runtime by the ServiceNow Super Intelligence Control Tower and enforced by NVIDIA OpenShell."

According to ServiceNow's release, Arc is part of ServiceNow EmployeeWorks and is "included at no additional cost with all Foundation, Advanced, and Prime SKUs." ServiceNow says "Mac support is live and Windows support is expected to roll out within weeks." It introduced Arc as Project Arc in an early preview on May 5, 2026.

The hardware layer comes later. ServiceNow says "AI Passport, including enforcement in NVIDIA OpenShell and NVIDIA BlueField DPUs, is expected to be available in a future release."

"As AI agents take on more autonomous tasks, governance and security have to extend to everything they touch," said Jon Sigler, executive vice president and general manager, AI Platform at ServiceNow. ZDNET also reported the launch, noting that Arc is available on Mac and that Windows support "is still on the way."

#How to get it

NVIDIA links its "now available" announcement to the Secure Agent Workspace Reference Design documentation. For Sentry, NVIDIA points developers to DOCA Argus, DOCA Flow and OVS-DOCA "for experimenting with out-of-band observability and infrastructure enforcement."

#How does it compare?

SDxCentral reported that OpenShell "debuted at Nvidia’s flagship GTC event in the Spring," and that the new design extends it to full-stack infrastructure. For other ways companies are putting controls around agents, see our coverage of Microsoft's insider-risk controls for AI agents and our guide to controlling what AI coding agents can reach on your Mac. More agent infrastructure coverage is in the Agent Platforms hub.

#What is still unclear?

  • Sentry date. NVIDIA says only "early next year."
  • Price. No source gives one for the architecture, OpenShell, Sentry or BlueField-4.
  • Production use. NVIDIA's release does not say which partners run the architecture in production. The ServiceNow Arc general availability it mentions is tied to OpenShell, not Sentry.
  • Performance baselines. NVIDIA gives no baseline for the 30x and 20% figures.