Dynamic pathways of information flow between distributed brain regions underlie the diversity of behavior. However, it remains unclear how neuronal activity in one area causally influences ongoing population activity in another, and how such interactions change over time. Here we introduce a causal approach to quantify cortical interactions by pairing simultaneous electrophysiological recordings with neural perturbations. We found that the influence primary visual cortex (V1) and higher visual area LM had on each other was surprisingly variable over time. Both feedforward and feedback pathways reliably affected different subpopulations of target neurons at different moments during processing of a visual stimulus, resulting in dynamically rotating communication dimensions between the two areas. The influence of feedback on V1 became even more dynamic when visual stimuli were behaviorally relevant and associated with a reward, impacting different subsets of V1 neurons within tens of milliseconds. Importantly, these fast changes in inter-areal influences were in stark contrast to, and could not be explained by, the much slower dynamics of activity in either area. To understand the function of dynamically rotating communication dimensions, we used a linear dynamical system to model the recurrent dynamics and external inputs that shape V1 activity. We found that the fast rotation of feedback communication dimensions momentarily aligned feedback influences with dynamical modes in V1 and with the visual input, creating a selective, transient time window for signal integration. Interestingly, only during this brief (<100 ms) time window, the feedback input to V1 was relevant for behavioral performance in a visual discrimination task. In summary, using a causal method for measuring long-range cortical interactions, we found that communication subspaces between visual areas are dynamically rotating. This rotation leads to momentary alignment of external inputs with V1 dynamics, consistent with the formation of transient windows for signal integration and sensory processing.