Agent scheduling algorithm has great influence on the efficiency of executing task in agent based network management. The existing algorithms don't involve the relationship among tasks, so that considerable network load and waiting time will be incurred when confronted with complicated task systems. To solve scheduling problem in network management, a scheduling framework that answers for the characteristic of network management is presented and a semi-online multi-agent scheduling algorithm based on dependences of sub-tasks is proposed. The competitive analysis and proof reveal that the semi-online scheduling algorithm is better than the existing full-online scheduling algorithm. The results of the tests of performance and scheduling time of SONL are consistent with the theoretical results. This scheduling algorithm is a new approach for dynamic agents scheduling in network management.