NVIDIA publishes OpenShell reference architecture for agents
NVIDIA has published a reference architecture that runs its open-source OpenShell agent runtime on Vera CPUs, and partner CoreWeave reports that sandboxes start more than three times faster than on x86 servers.
NVIDIA detailed a reference architecture for securing and governing long-running AI agents. It builds on OpenShell, the open-source runtime we covered at its 0.1.0 release. NVIDIA's argument is that "an agent cannot govern itself," so OpenShell treats every agent workload as untrusted for its entire run.
The runtime has four parts:
- Gateway: the control plane that manages sandbox lifecycles and policies.
- Supervisor: runs separately from the sandbox, checks outbound requests against policy and controls access to credentials.
- Sandbox: isolates the agent and limits its filesystem and process access.
- Policy prover: uses formal verification to check that a policy's permissions stay within a limit the organization sets.
Operators can extend it in two ways. Supervisor middleware can inspect, block or rewrite an agent's network traffic, and gateway interceptors can add governance rules for a specific deployment without changing the gateway. NVIDIA says identity, authorization and audit systems plug in at these enforcement points.
The architecture runs on Vera, the 88-core CPU NVIDIA has been pitching for agent fleets. NVIDIA claims Vera delivers up to 2x performance on agent workloads and 2x performance per watt compared with x86, along with up to 1.2 TB/s of LPDDR5X memory bandwidth.
The new numbers come from CoreWeave. It found sandboxes started more than 3x faster on Vera than on x86, and that a single rack supported more than 11,000 concurrent agent environments. On Terminal Bench, CoreWeave measured Vera at 1.8x faster than x86. The post does not say which x86 parts were used for comparison, and none of the results have been independently verified.