This thesis deals about human behaviors simulation in an urban context. We focus on the behavior believability (as judged as external observers). That is why our agent have some anticipatory skills, which allow them to use predictions about their near future. Our architecture is a hybrid one, which is very innovative because of its functioning with « high-level modules », which are seen as black-box from the rest of the architecture. Their number and intern modeling are completely free. This makes our architecture very modular and generic, and it is important because the urban simulation domain has many different applications (urbanism, video games, security, etc.), with different constraints. However, this genericity brings another problem, which is the integration of several heterogeneous behavior into the same decisional process. This issue is addressed thanks to a behavior composition mechanism. To conclude, we ensure the scaling up of our architecture with the creation of several levels of detail in the agents modeling.