NVIDIA OpenShell is a secure runtime for AI agents that governs multi-agent fleets in production, enforcing access control, credential isolation, and audit logging outside the agent, in the infrastructure layer beneath it. Enforcement uses Landlock, a Linux kernel mandatory access control mechanism, layered with network, process, and I/O confinement across six policy domains; every sandbox starts with zero permissions until a signed policy grants access. Credentials are scoped, injected at runtime, and refreshed rather than held by the agent. Sub-agents receive narrower, auditable authority than their parent, and denied actions convert into scoped approval workflows. Sandboxes reconnect and recover across process termination and model context limits. This lets security and platform teams move agents from pilot to production with policy-verified enforcement and a fleet-level audit trail. NVIDIA OpenShell runs unmodified across Claude Code, Codex, OpenCode, and custom agents, and deploys on Kubernetes, Red Hat OpenShift, and enterprise Linux.