Modern 5G Radio Access Networks increasingly rely on cloud-native infrastructures and O-RAN principles to enable flexible, programmable, and automated network management. However, operating such systems remains challenging: network intents must be translated into valid configurations, mapped to heterogeneous resources, and applied through strict APIs typically expressed as Kubernetes Custom Resource Definitions (CRDs). This process is error-prone, time-consuming, and difficult to scale, especially when configurations involve complex nested structures, interdependent resources, and dynamic updates. This thesis investigates a multi-agent approach to intent-driven RAN configuration management, combining the natural-language understanding capabilities of Large Language Models (LLMs) with deterministic schema validation and operator-controlled safety mechanisms.