In the context of ubiquitous computing, the application execution environment is made of heterogeneous devices belonging to different users (PC, PDA, mobile phone, etc.). These characteristics require to structure the application in a network of software units that are relatively independent but may interact with each other. In this thesis, we propose MaDcAr-Agent, a model of self-adaptive component-based agents provided with a framework dedicated to adaptation. This model is characterized by a meta level which contains an assembling engine in charge of automatic and dynamic adaptations according to the agent's context. The behavior of the meta level is ruled by two specified policies: the assembling policy that allows an agent to adapt itself because of contextual changes according to the available components and the content management policy that allows an agent to obtain the components it mostly needs, by using interactions with other agents. These specifications are made explicit and uncoupled from the agent's applicative behavior. The agent designer uses them to take into account the disturbance of the system from unpredictible and repeated changes, without having a detrimental effect on the agent's autonomy. To validate our approach, several experimentations have been conducted with this model. For instance, we describe a scenario involving mobile robots that must explore autonomously an unknown area.
