Why AI Agent Fleets Need a Kubernetes-Style Control Plane to Stay Safe

A developer building a multi-agent system found that a Kubernetes-inspired control plane was essential for managing AI agents safely in production environments. Through a series of deliberate stress tests, the system successfully blocked uncertified agents, detected swapped model identities, and killed a runaway cost-burning agent mid-run at its budget ceiling. A key insight emerged from one scenario where an agent had not been modified at all, yet quietly degraded in performance — a drift that only continuous benchmarking caught, not any rule-based gate. Emerging tools like kagent and agent-sandbox bring Kubernetes-native primitives such as CRDs, RBAC, and workload isolation to AI agents, but the author argues these address placement and isolation without enforcing ownership or proven capability. The core thesis is that admission control — deciding what is allowed to run and why — matters more than orchestration alone, and that the most dangerous agents are those that drift unnoticed rather than those that are visibly altered.