Adaptive Self-Improvement for Smarter Energy Systems Using Agentic Policy Search

Controlling energy systems usually involves manually designed policies for decision-making, which can be complex and time-consuming to develop. This process requires interdisciplinary collaboration among multiple domain experts, resulting in slow and inflexible adaptation to rapidly changing environments. Large Language Models (LLMs) offer a promising paradigm shift by integrating extensive contextual knowledge with the capability to generate structured, executable code. We present Agentic Policy Search (APS)—a novel hierarchical optimization framework in which LLMs act as autonomous agents that propose complete control logics, translate them into executable code, and iteratively improve them through direct system feedback. We apply APS to a residential energy system with PV, battery, demand, and dynamic electricity prices. Within just seven simulated days, the method yields a net profit of up to 6.20 € compared to the no-battery reference scenario (-10.70 €), nearly matching the global optimum of a perfectly informed linear program. By combining LLM-driven policy search with the generation of human-interpretable control logic, APS effectively bridges adaptability and traceability in energy management— while also offering a transferable framework for agentic optimization in other domains.