The last decades have seen the most important improvements of the safety on the road. But most of the accidents are still coming from the driver when a source of danger is not noticed or when the reaction is not correct during an unusual situation. Two ways of research are currently under investigations. Full autonomous vehicles are developed in order to avoid any mistake coming from the driver. Or the driver's capacity has to be improved with artificial devices. But neither the driver nor a technology can always guarantee full safety. Therefore a new direction of investigation must be taken by looking at an adaptive cooperation between both entities to estimate the most reliable vehicle command possible. This thesis tries to improve the reliability of the driving safety by mitigating the influence of both driver and intelligent control on the executive level. First of all, the vehicle architecture will be described. The use of transfer functions will lead to an optimal realization of the command and to a predictive vehicle model. Then the integration of several assistant systems into a multiagent software architecture is performed to define a reliable and flexible virtual driver. This unit will be able to compute the safety envelope of the vehicle and also the optimums of the different feasible maneuvers. The cooperation on the command level can be realized depending on their confidences and the driver's request. Therefore the driver's attention needs to be evaluated continuously to define the confidence value of the driver's command. Unfortunately these two entities may conflict with each other if one actor has a lack of knowledge. In such cases the situation analysis will be upgraded by sharing the knowledge from each one and give as feedback to the other.