Models of sociocultural evolution generally study the population dynamics of cultural traits given known biases in social learning. Cognitive agency, understood as the dynamics underlying a specific agent’s adoption of a given trait, is essentially irrelevant in this framework. This article argues that although implementing and instrumenting agency in computational models is fundamentally challenging, it is ultimately possible and would help us overcome major limitations in our understanding of sociocultural dynamics.Indeed, the behaviour of humans is not causally generated by a set of predefined behavioural laws, but by the situated activity of their cognitive architecture. Idealised models of biased transmission certainly help us understand specific features of population dynamics. However, they distract us from the deep intrication of the cognitive and ecological processes underlying sociocultural evolution, and erase their embodied, subjective nature.In line with the earlier “Thinking Through Other Minds” account of sociocultural evolution, this article highlights how the Active Inference framework can help us implement and instrument computational models that address these limitations. Such models would not only help ground our understanding of sociocultural evolution in the underlying cognitive dynamics, but also help solve (or frame) open questions in the study of ritual, relation between cultural transmission and innovation, as well as scales of cultural evolution.