At present, the primary control system in the field of rail transit is in the form of centralised control, which is based on the communication of the ground side system and the on-board side system. A central controller will play the role of information flow centre and has to afford a lot of computational loads when it needs to manage a large number of trains simultaneously. It is believed that with the development of artificial intelligence algorithms, the control system could be built based on a semi-decentralised multi-agent system (MAS). This paper proposes an innovative MAS system to enhance the decision-making of the rail transit control system. This proposed MAS includes several agents, e.g., train agents, station agents, and a central agent. A train agent can exchange information with other agents directly and make decisions based on the collected data. A case study is carried out to build a preliminary MAS for a rail transit control in Suzhou Metro. Compared with the centralised control system, this MAS can reduce the computational pressure of the central controller and improve the information exchange efficiency. The proposed method is expected to advance the thinking of how to achieve fully automated train control in the future.