Agentic AI Optimizes Policy-Driven Physical Layer Systems Long-Term

Network operators face a complex challenge in optimizing physical layer systems due to constantly changing policies, service requirements, and stringent real-time constraints. Traditional optimization methods, designed for fixed objectives, often prove ineffective in such dynamic environments. Researchers propose Agentic long-term performance optimization (Agentic-LTPO), a novel nested bilevel optimization framework that leverages agentic AI. In this framework, the upper level employs agentic AI to generate adaptive configurations for physical layer problems. These configurations are derived by translating evolving operator policies, environmental summaries, and historical experiences into structured lower-level optimization problems. The lower level then solves these updated problems to make real-time physical-layer decisions. Using cell-free MIMO beamforming as a practical use case, Agentic-LTPO was embodied with a multi-agent decision process and retrieval-augmented experience-based verification in the upper level, paired with a closed-form beamformer in the lower level. Experiments showed that Agentic-LTPO exhibited robust adaptability to dynamic operator policies and significantly improved the system's long-term performance by 57.2% compared to conventional approaches.