Nvidia launches OpenShell design for controlling AI agents
The architecture includes Sentry software, which Nvidia says can stop agents within milliseconds if they try to cross their permitted boundaries.
Nvidia has released its OpenShell Reference Architecture, a computing design intended to keep AI agents within defined operating limits on enterprise systems and edge machines. Announced at GTC Berlin, the architecture is available now, the company said.
The design assigns different tasks to Nvidia components: Vera Rubin systems run the models, and Vera CPUs run the software behind agents and their tools. OpenShell, an open-source runtime, establishes the limits on agents’ actions.
Nvidia Sentry software runs on BlueField-4 data processing units and monitors agents from outside the main system. According to Nvidia, an attempt to cross a software boundary triggers quarantine and shutdown within milliseconds. The company also said Sentry can increase agent capacity by up to 20%.
Nvidia listed Anthropic, OpenAI, Google, Meta and Cohere among providers whose agents can use the infrastructure. Anthropic said its integration of Claude Managed Agents with OpenShell lets companies restrict agent activity and log every action. CrowdStrike, Cisco, Okta and Zscaler are among security companies adding support.
BBVA and Santander are examining the technology for safety and governance, while SAP is integrating it into Joule Studio. ServiceNow has made its personal agent Arc generally available, using OpenShell to enforce limits on its actions.
The architecture supports Linux, Kubernetes and Windows, as well as AWS, Google Cloud and Microsoft Azure. Its edge coverage includes RTX Spark, DGX Spark and DGX Station systems. Nvidia did not disclose pricing.
