Multiple Weak Agents Defeat a Single Strong One: Integrating Large Language Models to Transform Weak Reinforcement Learning Agents into a Supreme Agent