Abstract: An implementation mechanism for agent’s behavior conventions is proposed in this paper. An equation used to estimate the expected efficiency of a convention mechanism is presented and an implementation architecture based on artificial neural network is developed. Having taken into consideration the difference of rights of control over certain conventions, this mechanism efficiently embodies the role played by normative mechanisms as the critical pivot of macro micro links in MAS. In addition, the mechanism presented here is dynamic, so that the evolution of conventions can be realized. Compared with relevant works, this work incorporates the social difference between agents and the coercive and evolutionary nature on the part of conventions.