Modeling automated user tasks based on agent-oriented approaches is a promising but challenging task. Personalized user agents have been investigated as a potential way of addressing this issue. Most of recent research work has focused on learning, eliciting and reasoning about user preferences and profiles. In this paper, our goal is to deal with the engineering of such systems, barely discussed in the literature. In this context, we present a high-level domain-specific model whose aim is to give users the power to customize and dynamically adapt their user agents. In addition, we propose a general architecture for developing user-customizable agent-based systems.