Robustness, through fault tolerance, is a property often put forward in order to advocate MAS. The question is: What is the first step to be fault tolerant? Obviously the answer is: to know faults. The claim of this paper is that existing fault classification suitable for distributed systems does not fit completely MAS needs because of autonomy, the main characteristic of their components. Actually autonomy is the very distinctive concept of agents and has unquestionable worthwhile properties. But do these properties have no compensation? After a short presentation of the fault classification which prevails in fault tolerance community, the paper will show that autonomy induces a need for significant extension to this classification. It will then make a special review of this extension and present some expectations with regard to the programing of fault tolerant MAS. Key words: Fault tolerance, MAS design, autonomy