This paper describes the progress of our work on autonomic fault-tolerant multi-agent systems. Our starting point is that new generation distributed applications are increasingly dynamic and open, e.g., new agents can join or leave, they can change roles or strategies and interdependences with other agents. Thus the traditional approach of configuring fault tolerance techniques (e.g., through replication) at design time may turn out inadequate. We propose to make autonomous the management of fault tolerance, more precisely by replication, along the course of the computation. Therefore, in a first step, we propose a framework for dynamic replication with some specific features (e.g., dynamic adaptation of replication strategies). We then propose several metrics for identifying dynamically what agents are more critical. In this paper, after surveying 2 metrics based on dependences and on roles, we detail our most recent metric, based on plans. We also discuss the issue of creation of replicas as a resource allocation problem.