Over the past few years, the prediction error minimization (PEM) framework has increasingly been gaining ground throughout the cognitive sciences. A key issue dividing proponents of PEM is how we should conceptualize the relation between brain, body and environment. Clark advocates a version of PEM which retains, at least to a certain extent, his prior commitments to Embodied Cognition and to the Extended Mind Hypothesis. Hohwy, by contrast, presents a sustained argument that PEM actually rules out at least some versions of Embodied and Extended cognition. The aim of this paper is to facilitate a constructive debate between these two competing alternatives by explicating the different theoretical motivations underlying them, and by homing in on the relevant issues that may help to adjudicate between them.