Peer interaction has been found to be conducive to learning in many settings. Knowledge co‐construction (KCC) has been proposed as one explanatory mechanism. However, KCC is a theoretical construct that is too abstract to guide the development of instructional software that can support peer interaction. In this study, we present an extensive analysis of a corpus of peer dialogs that we collected in the domain of introductory Computer Science. We show that the notion of task initiative shifts correlates with both KCC and learning. Speakers take task initiative when they contribute new content that advances problem solving and that is not invited by their partner; if initiative shifts between the partners, it indicates they both contribute to problem solving. We found that task initiative shifts occur more frequently within KCC episodes than outside. In addition, task initiative shifts within KCC episodes correlate with learning for low pre‐testers, and total task initiative shifts correlate with learning for high pre‐testers. As recognizing task initiative shifts does not require as much deep knowledge as recognizing KCC, task initiative shifts as an indicator of productive collaboration are potentially easier to model in instructional software that simulates a peer.