Imagine a team of specialized AI agents working together to debug code, learning from mistakes, and refining their approach with each iteration. This is the essence of RGD, a novel framework for enhancing code generation using multiple Large Language Models (LLMs). Traditional code generation with LLMs often hits roadblocks when dealing with complex tasks. The code might work for a few test cases but fail in unexpected ways when faced with real-world scenarios. This is where RGD comes in, introducing a collaborative debugging system inspired by how human programmers work. RGD employs three distinct LLM agents: the Guide, the Debugger, and the Feedback Agent. The Guide creates a strategic plan for code generation based on the task description. The Debugger writes the code, following the Guide’s instructions. And crucially, the Feedback Agent analyzes the results, pinpointing errors and suggesting improvements, just like a human debugger would. This isn’t just about generating code; it’s about building AI that understands why the code works or fails. RGD leverages a 'memory pool' of successful guides and task descriptions. This memory helps the Guide create more effective strategies, learning from past successes to improve future code generation. The Feedback Agent plays a crucial role by considering both failing and passing test cases, ensuring that fixing one bug doesn't inadvertently create another. By learning from both successes and failures, the system continuously refines its approach, getting closer to a perfect solution with each iteration. Experimental results show RGD significantly outperforms existing methods, particularly with complex tasks on the HumanEval, MBPP, and APPS datasets. The findings highlight RGD’s effectiveness in teaching LLMs to self-debug and adapt, paving the way for more robust and reliable AI-generated code in the future.