Clark’s (2013) prediction-action cognitive architecture predicts that active learners are more likely to learn. Even though this vision is widely spread in education, it has not been rigorously tested yet. This study investigated if agency while playing Mecanika, a serious educational game about Newtonian Physics knowledge, is beneficial to learning. Participants were 74 French-speaking undergraduate students with a novice background in Physics. Participants were paired and randomly designated as an active player or a passive watcher. Players played Mecanika for two hours while watchers were looking in real time on a separate screen a duplicate of the player’s screen. Dual-EEG was recorded to derive cognitive load and cognitive engagement metrics. Before and after the experiment, participants filled a Physics knowledge questionnaire, the Force Concept Inventory (Hestenes et al. 1992), where each of the five multiple choices for every item was either representing a scientific conception (good answer), and common misconceptions or fillers (wrong answers). Transitions occurring from pretest to posttest according to those three types of answers were analyzed. One unanticipated finding was that agency is less beneficial to learning than passively watching: watchers are more likely than players to transition from a wrong answer (either misconception of filler) to a good answer (scientific conception) between the pretest and the posttest. Statistical analyses of continuous psychophysiological measures of cognitive load and cognitive engagement didn’t show differences between players and watchers. Results point to a need for further research aiming at finer-grained decompositions of the performance and the learning context.