A detailed walkthrough of implementing a multi-agent supervisor pattern on OpenShift using OpenShell sandboxes to enforce least-privilege isolation for AI agents. The pattern addresses the security risks of a single AI agent holding credentials for multiple systems (Prometheus, Loki, Confluence) by splitting work into specialized agents, each running in its own sandbox with kernel-enforced network policies (Landlock, seccomp). The architecture uses Claude as the orchestrating brain in Anthropic's cloud, while actual data access and reasoning happen inside per-agent OpenShell sandboxes on the cluster. Each sandbox gets a scoped network policy allowing only the specific binary to reach specific endpoints, limiting blast radius if any agent is compromised via prompt injection. The post includes a full incident-response demo showing parallel agent dispatch, per-agent policy enforcement, and result synthesis without raw PII ever leaving the cluster.
