While the contemporary achievements of AI and robotics are indisputable, the issue of autonomy for artificial agents still looms ahead despite technological progress and rich conceptual debates. Drawing on recent theoretical propositions from the enactive approach on autonomy, we first highlight several limitations of what we call an identity-based model. Through the study of four real-life cases, we then not only argue that autonomy cannot be conflated with behavioral self-maintenance or organizational closure, but that it can sometimes violate these conditions. Finally, we propose a naturalistic activity-based model of autonomous agents that emphasizes the importance of norm-establishing processes distributed across an intricacy of milieus.