Real-time animation of virtual humans requires a dedicated architecture for the integration of different motion control techniques running into so-called actions. In this paper, we describe a software architecture called AGENTlib for the management of action combination. Considered actions exploit various techniques from keyframe sequence playback to inverse kinematics and motion capture. Two major requirements have to be enforced from the end user viewpoint: first, that multiple motion controllers can control simultaneously some parts or whole of the virtual human, and second, that successive actions result in a smooth motion flow