Virtual humans in large environments A virtual human is a piece of software that imitates behaviour of a human in a virtual world and that is equipped with a virtual body, visualized by a graphical viewer. A complex virtual human carries out more complicated tasks than just walking, object grasping or chatting in Elisa-like manner. Most research applications focused on "artificial intelligence" of virtual humans exploit either small environments with a few complex actors, or a large environment with many actors that are, however, not complex (but carry out walking, crowding etc.). Anyone who aims at developing a virtual reality application featuring a large artificial world inhabited by complex virtual humans, e.g. a role-playing computer game or a large storytelling application, will face at least two issues that are delitescent if one is exploiting a small environment or simple actors only. They are the problem with simulation speed and the problem with managing extensions, both during development and after release. First, it is almost impossible to run tens of complex actors on a single PC because of limited computational and memory resources. Second, because of design purposes, inserting new objects and actions should be allowed anytime, both during the development and after the release without using any machine-learning technique. We have developed a framework that copes with these issues. The solution is based on augmentation of the level-of-detail AI technique, theories of affordances by J. J. Gibson, and practical reasoning by M. E. Bratman, and action selection mechanism POSH of J. Bryson. The framework will be presented in the talk and its main principles will be discussed.