Modern training typically incorporates real-world training applications. Augmented Reality (AR) technologies support this by overlaying virtual objects in real-world 3-Dimensional (3D) space. However, integrating instruction into AR is challenging because of technological and educational considerations. One reason is a lack of architecture for supporting Intelligent Tutoring Systems (ITSs) in AR training domains. We present a novel modular agent-based Distributed Augmented Reality Training (DART) architecture for ITSs to address two key AR challenges: (1) a decoupling of the display and tracking components and (2) support for modularity. Modular agents communicate with each other over a network, allowing them to be easily swapped out and replaced to support differing needs. Our motivation is driven by the fact that AR technologies are vary considerably and an ITS architecture would need to be flexible enough to support these requirements. Finally, we believe that our novel architecture will appeal to practical designers of ITSs and to the more theoretical educators who wish to use such systems to simulate and broaden research in the distributed cognitive educational theories.