Can you learn better by doing something yourself (DIY) or by watching somebody else do it? We present a new approach to examine this perennial question in research on learning and instruction. In a science learning task, children aged 5 to 7 years (N = 95) either generated predictions themselves (active condition) or observed the predictions of a fictitious other child (yoked condition) before seeing the outcome. Unlike previous yoked designs, we first emulated responses from a Bayesian learner. Critically, these responses were then individually matched to each yoked child given their unique prior beliefs at the start of the experiment. This novel approach allowed us to discern the effects of DIY on conceptual learning much more clearly than before. We found that actively generating predictions led to deeper conceptual understanding than observing another's matched predictions, and that this advantage of DIY was associated with an increased experience of agency.