The columnar hypothesis holds that neurons in a single cortical column process similar types of information. Here we show, however, that the way mouse primary visual cortex integrates visual and non-visual information differs fundamentally between layers. We used large-scale two-photon calcium imaging and Neuropixels recordings to compare population activity across layers in awake mice. Layer 2/3 activity is more strongly modulated by visual stimuli, whereas layer 5 activity is more strongly modulated by the animal's movement. Furthermore, movement has opposite effects on population synchrony in the two layers: it desynchronizes the spontaneous oscillations of layer 2/3 while synchronizing layer 5 activity. The geometry of population activity also differs, with spontaneous activity in layer 2/3 being lower-dimensional and more dominated by a single coding dimension than in layer 5, particularly during rest. These results suggest a division of labor, where superficial layers generate a robust sensory representation for downstream cortical processing, while deep layers integrate sensory signals with behavioral context to inform motor output.