The author past studies on production planning include agent systems where agents act independently under the real-time control of Reasoning to Anticipate the Future (RAF) in order to realize an autonomous decentralized flexible manufacturing systems (AD-FMS) and achieve high production efficiency. However, RAF did not solve the automated guided vehicles (AGVs) colliding problem. This paper describes the method of cooperation by knowledge exchange in AGVs moving autonomously under the real-time control of RAF in AD-FMSs to avoid AGVs collisions. The method gives the AGV an individual knowledge called AGV-knowledge, and by the exchange of which, each AGV can avoid collisions. This method does not use the conventional control by a host computer but applies communication among AGVs. Head-on collisions were prevented by applying this method to 9 types of FMSs constructed in a computer.