How a user interacts with smartphone apps depends on the system recommendations (actions) in a dynamic way. While a SMART (Sequential Multiple Assignment Randomized Trial) allows for unbiased evaluation of the optimal action at each decision point, its use in practice could be limited due to the lack of guidance on how such a trial should be designed and implemented. In this talk, I will discuss design issues of a SMART in an application where we build an app recommender system to maximize user engagement in a suite of health apps for depression and anxiety disorders. Specifically, I will present some recent findings regarding sample size determination and adaptation of a SMART. The goal of our work is to enhance the practicality of SMART-type designs for improving the experience of health app users.