Few early intervention efforts have improved the representation of women in computer science and engineering (CSE) disciplines, but pair programming has shown promise for reducing gender differences among college students. The current study is the first to examine this promising practice in middle school. In an effort to better understand what pair programming looks like, we describe an observational study of middle school girls. We coded audiotape transcripts to show the kinds of interactions that appear to promote or undermine effective problem solving. The findings are interpreted in terms of how to promote the kinds of interactions that make it more likely that middle school students will persist in the kind of problem solving that will prepare them for further CSE coursework.