Despite the fact that autonomy is a central notion in most agents' definitions and it is by itself a subject of research in the multi-agent field, there is not yet a commonly agreed definition of it. This paper attempts to classify the different forms of autonomy proposed in the literature by using the Vowels approach. After underlining some of the most relevant properties of autonomy, the classification is used to propose a comprehensive definition of autonomy. Moreover, we identify two different perspectives on autonomy, namely external and internal, and we show that the later can be used to build an integrated analysis grid of agent architectures for multi-agent systems.