Behaviour-tracking technology has been used for decades in SLA research on focused practice with an eye toward elucidating the nature of L2 automatization (e.g. DeKeyser, 1997; Robinson, 1997). This involves longitudinally capturing learners' judgments or linguistic production along with their response times in order to investigate how specific skills become automatic over time. However, previous research in this area has been conducted mostly in laboratories (i.e., in vitro), sometimes with artificial languages, thereby compromising ecological validity of the findings. Building on this work, this article reports on a one-month study in which learners' (N = 126) behaviour was tracked while they practised two constructions of English grammar (varying in complexity) using mini-games that involved some time pressure and were embedded in meaning-focused reading and discussion activities in class. Feedback was randomly varied between participants. Multilevel statistical analyses of accuracy and response time suggest that practice helped to develop automaticity, and that rule complexity and metalinguistic feedback played a role. The methodological innovation of this study consists of the application of in vitro experimental research techniques in in vivo L2 learning contexts and of the use of statistical mixed effects models to account for the complexity of real-life tracking data.