Abstract: Conflict-free path planning is critical for the efficient and safe operation of automated guided vehicle (AGV) logistics systems in intelligent manufacturing workshops. Traditional centralized decision-making approaches for solving multi-AGV conflict-free path planning (MCPP) problems face challenges such as insufficient environmental adaptability and flexibility. Therefore, a distributed decision-making approach is applied to accommodate the complex and dynamic logistics requirements in intelligent manufacturing workshops. Firstly, based on the grid map-based workshop logistics transport environment, the MCPP problem is described, potential AGV conflict types are analyzed, and a conflict reduction strategy is designed. Then, the MCPP problem model is transformed into a partially observable Markov decision process corresponding to reinforcement learning (RL), and the RL elements, including the state space, action space, and reward function, are designed. Furthermore, an improved QMIX is employed to solve the MCPP problem. Finally, experimental studies are conducted to verify the feasibility and effectiveness of the proposed model and method through comparisons with the spatiotemporal A^* algorithm, conflict-based search algorithm, game theory method, value decomposition networks (VDN), and the classic QMIX algorithm.