Nvidia introduced the OpenShell Reference Architecture at GTC Berlin, a full-stack design for monitoring and containing autonomous AI agents. The system utilizes Vera CPUs for agent operations, Rubin systems for model execution, and BlueField-4 chips to manage security through the Sentry framework.
The OpenShell runtime treats agent workloads as untrusted, utilizing a sandbox, a gateway for policy management, a supervisor for credential handling, and a policy prover for verification. The BlueField-4 data processing unit inspects traffic between agents and models to ensure security independently of the agent's own reporting.
Nvidia bundled these technologies into the Open Agent Safety Platform on September 28, supported by over 100 launch partners. The architecture is designed to support agents from providers including OpenAI, Anthropic, Google, and Meta, with security vendors like CrowdStrike and Palo Alto Networks integrating support.
Enterprises such as BBVA and Santander are exploring the framework for governance, while ServiceNow has launched its Arc agent with runtime actions enforced by OpenShell. While the core architecture is available now, the Sentry framework is scheduled for release early next year, with additional resources currently in testing.
