Many Virtual Reality applications, such as training applications, urban design or gaming are based on a rich semantic description of the environment. This paper describes our proposal for a representation model of semantic virtual worlds. Our model called SVHsIEVs1 should provide a consistent representation of the following aspects: the simulated environment, its structure and the behavior of its elements, and the knowledge items using ontology, interactions and tasks that virtual humans can perform in the environment. Our first contribution is to show the influence of adding semantic information (contextual attribute and relationships between their concepts) to virtual objects and using this influence to define and redefine interaction with the environment. Our second contribution is to use in our model the Virtual Human based on two technics: querying and guidelines. Our third contribution is to use these semantic information in management of tasks for each virtual object; we proposed to define each task by a set of attributes and relationships, which determines the links between attributes in tasks, and links between other tasks. This paper presents a multi-agent architecture oriented to Virtual Human in Semantic virtual environments. The architecture has been successfully tested in 3D dynamic environments with navigation in virtual urban environment.