Abstract: With the increase of source and load types, the real-time optimal scheduling of AC/DC hybrid distribution network becomes more and more difficult. Reinforcement learning agent has the mode of "offline training, online application", which can effectively solve such real-time optimization problems. However, the centralized optimal scheduling method proposed at present has high requirements for computing equipment and cannot guarantee the privacy of station data. Therefore, this paper proposes a multi-agent based optimal scheduling method for medium and low voltage AC/DC hybrid distribution networks. Firstly, based on the power flow calculation method of Distflow, the mathematical model of optimal dispatching of low-voltage AC/DC hybrid distribution network is established. Secondly, the node voltage and power flow distribution information of each station area is taken as the state space, the operation of equipment that can be optimally dispatched in each station area is taken as the action space,the minimum network loss and node voltage deviation in each station area is taken as the reward function, a Markov game model is established, which is solved by adopting the "Actor-Critic" reinforcement learning algorithm, thus realize the multi-equipment distributed coordinated control of AC/DC hybrid distribution network based on multi-agents. Finally, taking a 21-node medium and low voltage AC/DC hybrid distribution network as an example, the feasibility and effectiveness of this method in controlling the voltage over-limit problem and reducing network loss are verified.