OpenShell uses four components to control agents: a Gateway that manages sandbox lifecycles and policies, a Supervisor that checks outbound requests and mediates credential access from outside the sandbox, a Sandbox that isolates the agent's filesystem and process access, and a Policy Prover that formally checks permissions against an organization-defined ceiling.
Extension points let teams add their own governance. Supervisor middleware can inspect, block or modify network content, and gateway interceptors add deployment-specific rules to control-plane operations without changing the gateway. In the reference design, OpenShell runs on NVIDIA Vera CPUs, which NVIDIA says deliver up to 2x performance on critical agentic workloads versus x86.
CoreWeave measured Vera against a leading x86 CPU and reported sandbox startup more than 3x faster and agentic loop completion 1.8x faster, with a single rack supporting more than 11,000 concurrent agent environments. For deployments that need more protection, NVIDIA Sentry running on BlueField-4 monitors agent behavior from a separate domain and can quarantine an agent within milliseconds of a detected violation.
Why it matters if you build agents
Teams building agents can enforce policy on network traffic, credentials and filesystem access outside the agent process instead of relying on the agent to restrain itself. The Supervisor middleware and Policy Prover offer places to add checks, though identity, authorization and auditing still need to be integrated separately.
The sandbox density and speed figures come from NVIDIA and CoreWeave, so teams should benchmark their own workloads before sizing hardware.
Get stories like this every weekday
One short email with the day's agent tooling news.