In a typical dextrous manipulation task, a goal configuration is reached through a sequence of continuous motions. Most often, a motion plan is computed offline and subsequently used as a reference trajectory for a feedback controller. We present an alternative approach that only relies on feedback, no motion planning is necessary. Different feedback controllers are constructed and composed in such a way that the object moves toward the goal configuration. Switches between the controllers are not planned ahead, they result from the feedback itself. Discrete features such as finger gait are therefore generated on-line. The approach is based on our previous results on stabilization of hybrid systems.