Problem

Traditional radio access network (RAN) management is complex and requires significant human intervention. The advent of agentic artificial intelligence (AI) and Large Language Models (LLMs) offers a path towards autonomous RANs, where high-level operator intents can be translated into network optimizations. However, existing agentic AI solutions for Open RAN (O-RAN) typically address simple intents with independent agents, overlooking the complexities of inter-agent coordination for more sophisticated goals. Furthermore, deploying individual large LLMs for each agent presents a significant scalability challenge due to high memory consumption.

Method

This paper introduces an agentic AI framework designed for intent-driven optimization within cell-free O-RAN. The framework consists of multiple LLM-based agents that operate at various timescales and collaborate through standardized interfaces. A Supervisor agent, located in the non-RT RIC, is responsible for interpreting operator intents and translating them into a specific objective function and minimum rate requirements. A User Weighting agent in the near-RT RIC leverages prior experience from a memory module to determine user priority weights for precoding. In scenarios requiring energy saving, an O-RU Management agent utilizes a multi-agent Deep Reinforcement Learning (MAPPO) algorithm to decide which Open Radio Units (O-RUs) should be active. A Monitoring agent continuously tracks user data rates and orchestrates the User Weighting and O-RU Management agents to ensure that all minimum rate requirements are met, adjusting parameters as needed. To enhance convergence speed, a retrieval-augmented coefficient tuning method stores and reuses previously learned coefficient sets. For scalability, the framework employs Parameter-Efficient Fine-Tuning (PEFT) through Quantized Low-Rank Adaptation (QLoRA) adapters on a shared lightweight LLM in the near-RT RIC, significantly reducing memory overhead while maintaining agent specialization.

Results

Simulation results demonstrated the effectiveness of the proposed agentic AI framework, particularly in energy-saving mode. The framework achieved comparable performance with both 7B and 14B parameter Qwen models, outperforming baseline schemes by reducing the number of active O-RUs by up to 41.93%. In contrast, a DRL+gradient ascent baseline showed instability due to uncoordinated agent updates. A crucial finding was the substantial reduction in memory usage; by combining quantization and QLoRA, the framework decreased the overall memory footprint by approximately 92% compared to deploying separate, full-precision LLM agents. The retrieval-augmented tuning method proved beneficial in accelerating the convergence of user priority weights and violation penalty coefficients by leveraging stored historical data. The dynamic coordination between agents was also shown to effectively adjust network parameters to meet changing operator intents and maintain service quality.

Implications

This research presents a significant step towards fully autonomous, intent-driven optimization in cell-free O-RAN architectures. By enabling complex inter-agent coordination and addressing scalability challenges through PEFT techniques like QLoRA, the framework offers a robust and efficient solution for future wireless networks. The ability to translate natural language intents into concrete network actions, coupled with demonstrated improvements in energy efficiency and memory utilization, highlights the practical potential of agentic AI in evolving RAN management. The work opens avenues for future enhancements, including the integration of additional agents for tasks suchs as resource block allocation and channel estimation, further expanding the scope of autonomous network control.