Simulations were completed using artificial reading “agents” that are subject to known physiological (e.g., limited visual acuity) and psychological (e.g., limited attention) constraints and capable of learning to move their eyes and allocate attention to read as efficiently as possible. These simulations indicate that agents learn when and where to move their eyes to attain maximal reading efficiency, generalize this behavior from training sentences to novel test sentences, and use word length to predict word-identification times and thereby make optimal decisions about when to initiate saccadic programming—even if word length is only moderately predictive of word-identification times. These results suggest that humans may exploit even modestly informative cues in learning to decide when to move their eyes during reading.