This article investigates infant comprehension of filler-gap dependencies. Three experiments probe 15- and 20-month-olds' comprehension of two filler-gap dependencies: "wh"-questions and relative clauses. Experiment 1 shows that both age groups appear to comprehend "wh"-questions. Experiment 2 shows that only the younger infants appear to comprehend relative clauses, while Experiment 3 shows that when parsing demands are reduced, older children can comprehend them as well. We argue that this discontinuous pattern follows from an offset in the development of grammatical knowledge and the deployment mechanisms for using that knowledge in real time. Fifteen-month-olds, we argue, lack the grammatical representation of filler-gap dependencies but are able to achieve correct performance in the task by using argument structure information. Twenty-month-olds do represent filler-gap dependencies but are inefficient in deploying those representations in real time.