Real-time models of learning have gained considerable interest in the study of learning and development. In particular, Temporal Difference (TD) learning (Sutton & Barto, 1990) has received a great deal of attention as a model of animal learning, a model of activity of dopamine neurons in the primate midbrain, and as a control architecture for autonomous agents. TD learning constitutes a real-time generalization of the Rescorla-Wagner model of learning (Sutton & Barto, 1990; Rescorla & Wagner, 1972). As such, explains certain phenomena in much the same manner, and can account for a wide range of animal learning results. While other models of learning based on theories of eligibility or attention (Mackintosh, 1975; Kruschke, 2001) can better account for certain empirical findings (e.g., blocking), little work has been done to develop real-time extensions to these models. In this paper, a real-time model of attention is proposed, building on previous real-time models of reinforcement and trial-level models of attention. The model incorporates aspects both of TD learning and connectionist models of attention (Kruschke, 2001) and explains findings from human and animal research in which attentional mechanisms are implicated. Additionally, the development and function of the model is studied in an embodied context in which it is implemented as the control architecture for autonomous agents behaving in simplified environments. The real-time model of attention is found to produce adaptive behavior in a variety of environments in which previous real-time models do not.