The paper introduces the Multi-Agent Conversational Recommender System (MACRS), an LLM-only multi-agent system designed to improve conversational recommendation. MACRS addresses the challenges of controlling dialogue flow and leveraging user feedback in conversational recommender systems (CRS). The system consists of two main modules: a multi-agent act planning framework and a user feedback-aware reflection mechanism. 1. **Multi-Agent Act Planning Framework**: This module includes three responder agents (asking, chit-chatting, and recommending) and one planner agent. The responder agents generate candidate responses based on different dialogue acts, while the planner agent selects the most appropriate response. The planner agent also considers user feedback to adjust the dialogue act plan. 2. **User Feedback-Aware Reflection Mechanism**: This mechanism leverages user feedback at two levels—information-level and strategy-level—to optimize the system's performance. Information-level reflection infers user preferences from feedback and summarizes them into user profiles. Strategy-level reflection adjusts the dialogue act plan based on feedback, providing suggestions and corrective experiences. Experiments on the MovieLens dataset show that MACRS outperforms existing CRS methods in terms of recommendation accuracy and user preference collection. The system demonstrates better dialogue flow control, more engaging interactions, and improved user satisfaction, particularly for low-popularity items. The case study further illustrates how MACRS effectively captures and utilizes implicit user preferences, leading to more accurate recommendations.The paper introduces the Multi-Agent Conversational Recommender System (MACRS), an LLM-only multi-agent system designed to improve conversational recommendation. MACRS addresses the challenges of controlling dialogue flow and leveraging user feedback in conversational recommender systems (CRS). The system consists of two main modules: a multi-agent act planning framework and a user feedback-aware reflection mechanism. 1. **Multi-Agent Act Planning Framework**: This module includes three responder agents (asking, chit-chatting, and recommending) and one planner agent. The responder agents generate candidate responses based on different dialogue acts, while the planner agent selects the most appropriate response. The planner agent also considers user feedback to adjust the dialogue act plan. 2. **User Feedback-Aware Reflection Mechanism**: This mechanism leverages user feedback at two levels—information-level and strategy-level—to optimize the system's performance. Information-level reflection infers user preferences from feedback and summarizes them into user profiles. Strategy-level reflection adjusts the dialogue act plan based on feedback, providing suggestions and corrective experiences. Experiments on the MovieLens dataset show that MACRS outperforms existing CRS methods in terms of recommendation accuracy and user preference collection. The system demonstrates better dialogue flow control, more engaging interactions, and improved user satisfaction, particularly for low-popularity items. The case study further illustrates how MACRS effectively captures and utilizes implicit user preferences, leading to more accurate recommendations.