This paper proposes a Multi-Agent Conversational Recommender System (MACRS), an LLM-only system that can efficiently plan and dynamically refine its dialogue and recommendations. MACRS consists of two essential modules: (1) multi-agent act planning, which includes three responder agents and one planner agent, and (2) user feedback-aware reflection, a dynamic optimization mechanism based on the LLM. The first module plans dialogue acts for each turn, while the second module uses user feedback to adjust the dialogue act planning and generate higher-level user information. Extensive experiments on Movielens demonstrate the effectiveness of MACRS in recommendation and user preferences collection. MACRS addresses the challenges of controlling dialogue flow and incorporating user feedback in conversational recommendation systems. The multi-agent act planning module uses four LLM-based agents to generate various candidate responses and select the most appropriate one. The user feedback-aware reflection module leverages user feedback to adjust the dialogue act planning and generate higher-level user information. The experiments show that MACRS outperforms existing methods in recommendation accuracy and user preference collection. The paper also discusses related work, including LLM-based autonomous agents and conversational recommender systems. It highlights the limitations of existing methods and proposes MACRS as a solution. The methodology includes an overview of the system, multi-agent act planning, and user feedback-aware reflection. The experiments evaluate the performance of MACRS in terms of success rate, hit ratio, and average turns. The results show that MACRS outperforms existing methods in all metrics. The paper concludes that MACRS can significantly improve the performance of conversational recommendation systems and provide a better user experience.This paper proposes a Multi-Agent Conversational Recommender System (MACRS), an LLM-only system that can efficiently plan and dynamically refine its dialogue and recommendations. MACRS consists of two essential modules: (1) multi-agent act planning, which includes three responder agents and one planner agent, and (2) user feedback-aware reflection, a dynamic optimization mechanism based on the LLM. The first module plans dialogue acts for each turn, while the second module uses user feedback to adjust the dialogue act planning and generate higher-level user information. Extensive experiments on Movielens demonstrate the effectiveness of MACRS in recommendation and user preferences collection. MACRS addresses the challenges of controlling dialogue flow and incorporating user feedback in conversational recommendation systems. The multi-agent act planning module uses four LLM-based agents to generate various candidate responses and select the most appropriate one. The user feedback-aware reflection module leverages user feedback to adjust the dialogue act planning and generate higher-level user information. The experiments show that MACRS outperforms existing methods in recommendation accuracy and user preference collection. The paper also discusses related work, including LLM-based autonomous agents and conversational recommender systems. It highlights the limitations of existing methods and proposes MACRS as a solution. The methodology includes an overview of the system, multi-agent act planning, and user feedback-aware reflection. The experiments evaluate the performance of MACRS in terms of success rate, hit ratio, and average turns. The results show that MACRS outperforms existing methods in all metrics. The paper concludes that MACRS can significantly improve the performance of conversational recommendation systems and provide a better user experience.