A social network can be modelled by a multi-agent system, in which the interaction among agents is represented as a service transaction process. In this paper, we present a service-based agent network to simulate and study the dynamics of social networks. In the network, the profiles of agents and service-based interactions are defined deliberately. Autonomy is emphasized as the ability of agents to manage their behaviors according to the local environment and their profiles. The experimental results reveal that network performance, network topology and the profiles of agents all evolve along with local behaviors. The over-shoot phenomenon in the evolution of network is discovered and analyzed. The discoveries are meaningful for understanding the relationship between network dynamics and local behaviors.