This work focuses on the engineering of software replication techniques for distributed cooperative applications designed as multi-agent systems. Such applications are often very dynamic: e.g., new agents can join or leave, they can change roles or strategies. Also, the relative importances of agents may evolve during the course of computation and cooperation, as opposed to traditional static approaches of replication, e.g., for data bases, where critical servers may be identified at design time. Thus, we need to dynamically and automatically identify the most critical agents and to adapt their replication strategies (e.g., active or passive, number of replicas), in order to maximize their reliability and their availability. An important issue is then: what kind of information could be used to estimate which agents are most critical agents? In this paper, we first introduce our approach and prototype architecture for adaptive replication. Then, we discuss various kinds of information and strategies to estimate criticality of agents: dynamic dependences, roles, and plans. Some preliminary measurements and future directions are also presented.