CNCF: Kubernetes as the Operational Foundation for Agentic AI — Lessons from a Multi-Agent Security Platform

CNCF (Cloud Native Computing Foundation) published a case study in which a security architect from Orange Innovation explains why agentic AI belongs on cloud-native foundations.

Each Agent as a Separate Kubernetes Deployment

In the described multi-agent security platform, each AI agent is a standalone Kubernetes deployment, not part of a single monolithic service. This isolation means agents can be scaled, monitored, and constrained independently, just like microservices. The approach is the opposite of the common practice where the entire agentic system runs as one application, which makes control and resilience harder to achieve.

Policy-as-Code and a Monitored Model

Agent security decisions are governed through policy-as-code via Open Policy Agent (OPA) — a tool that expresses rules in code rather than manual configuration. Crucially, large language model calls are gated by classical anomaly detection: the LLM is not the last line of defense, but a component whose actions are verified by proven security methods before execution.

What Agentic AI Inherits from Cloud-Native

The author’s thesis is that agentic AI inherits all the operational problems cloud-native has already solved — observability, isolation, access management, delivery. Cloud-native primitives such as cert-manager, Cilium, and GitOps, it is argued, transfer directly to agentic workloads. The message is practical: teams do not need to invent a new operational discipline for agents, but simply apply the one the Kubernetes ecosystem already has.