In order to reliably support various services using multi-agent system, the faults in the system need to be tolerated. While various schemes have been proposed for fault tolerant MAS, the replication approach provides the basis of short recovery latency, less intrusive with respect to the execution time, and high scalability. In this paper we propose a new replication scheme making a decision on either replicating data or replacing the failed agent according to the current status of the agents. It thus greatly reduces the overhead of the fault tolerance mechanism by minimizing the amount of replication of the agents. The effectiveness of the proposed approach is verified using the agent-based hospital information provisioning system implemented on the agent platform developed by the authors. The experiment displays that the proposed approach significantly outperforms the existing approach in terms of recovery latency.