Recent advances in Large Language Models (LLMs) have significantly reshaped software engineering workflows, particularly in Automated Program Repair (APR). Despite their potential, LLMbased repair techniques often exhibit unstable behavior, limited generalization and strong prompt sensitivity. This proposal investigates a research direction that unifies Reinforcement Learning (RL), cognitive agents and self-adaptive architectures to autonomously optimize prompts for LLM-based patch generation. We hypothesize that treating prompt design as an adaptive optimization problem can significantly improve stability and correctness. By embedding prompt refinement in a continuous feedback loop, and enabling agents to maintain persistent beliefs extracted from repositories and execution traces, the repair process becomes more stable, explainable and effective.