Abstract Studying the early stages of transient events provides crucial information about the fundamental physical processes in cataclysmic variables (CVs). However, determining an appropriate observation mode immediately after the discovery of a new transient presents challenges due to significant uncertainties regarding its nature. We developed a framework designed for autonomous decision-making in prompt follow-up observations of CVs using the Kanata $1.5$ m telescope. The system, named Smart Kanata, first estimates the class probabilities of variable star types using a generative model. It then selects the optimal observation mode from three possible options based on the mutual information calculated from the class probabilities. We have operated the system for ${\sim}300$ d and obtained 21 samples, among which automated observations were successfully performed for a nova and a microlensing event. In the time-series spectra of the nova V4370 Oph, we detected a rapid deepening of the absorption component of the H$\alpha$ line. These initial results demonstrate the capability of Smart Kanata in facilitating rapid observations and improving our understanding of outbursts and eruptions of CVs and other galactic transients.