Distributed cooperative applications are now open, very dynamic and large scale. Thus, they are increasingly designed as Multi-Agent Systems (MAS).This evolution creates new challenges to the traditional approaches of fault-tolerance that mostly relies on centralized and offline solution. In this paper, we propose an adaptive replication-based preventive approach designed for large-scale MAS. Our solution stands on a negotiation protocol to dynamically and transparently adapt the agent replication strategy (number of replicas and their location) to the system's state. This protocol provides a decentralized resource allocation solution which uses local decisions to guarantee the global MAS reliability.