In this study, we investigate how to make takeover in conditionally automated vehicles safer optimizing the takeover request modalities. Using Machine Learning algorithms, we created a smart agent recommending the best modalities combination to use (haptic-visual, auditory-visual and haptic-auditory-visual) to convey a takeover request. This agent considers the driver state before recommending modalities, as well as the weather condition. The proposed agent is able to predict takeover quality better than the baseline by 56.5% (baseline MSE : 0.0600, our agent MSE : 0.0261) and use different recommendations based on the situation. Modality impact on takeover quality was shown to have a mean score of 4.95% (standard deviation: 2.7%). Evaluation of the agent gain on takeover quality compared to standard takeover request design is currently ongoing.