Designing a controller to supervise an ambient application is a complex task. Any change in the system composition or end-users needs involves re-performing the whole design process. Giving to each device the ability to self-adapt to both end-users and system dynamic is then an interesting challenge. This article contributes to this challenge by proposing an approach named Extreme Sensitive Robotic where the design is not guided by finality but by the functionalities provided. One functionality is then seen as an autonomous system, which can self-adapt to what it perceives from its environment (including human activity). We present ALEX, the first system built upon the Extreme Sensitive paradigm, a multi-agent system that learns to control one functionality in interaction with its environment from demonstrations performed by an end-user. We study through an evolutive experimentation how the combination of Extreme Sensitive Robotic paradigm and ALEX eases the maintenance and evolution of ambient systems. New sensors and effectors can be dynamically integrated in the system without requiring any action on the pre-existing components.