The invention discloses a multi-agent optimization control device and method for an automatic parking system. The device comprises a decision agent, area agents and execution agents. The decision agent continuously receives requests from all the area agents, performs real-time analytical calculation, sends instructions to all the area agents according to the overall operation time sequence of the automatic parking system and the current operating stage of the automatic parking system, and coordinates and distributes the operating time sequences of all the area agents. All the area agents continuously collect and analyze data information transmitted from the execution agents, and compare and analyze the data information and data stored in an application knowledge base so as to judge whether the environment where all the execution agents are located is normal or not and send control instructions to all the execution agents. The execution agents are responsible for collecting information of the environment around a vehicle, and the information is sent to the area agents through a data information exchange module.