In recent years, Internet of Things and Cyber-Physical Systems have became a major asset, shaping almost all aspects of human life. Connected objects host physical (e.g., pollution sensing) or software (e.g., positioning on a map, itinerary planning) services. Thanks to artificial intelligence, these basic services can be automatically composed to offer composite smart services (e.g., guidance in a city) tailored to the user in her or his current situation. However, mobility and unpredictability of user needs and service’s availability make the ambient environment highly unstable. In such contexts, providing working and relevant composite services is a challenging task: it demands middleware tools working despite uncertainty while maintaining the user in the loop. Therefore, we propose an innovative approach called “opportunistic composition”: using a bottom-up approach, composite services are built, from available basic services and human-machine interaction fragments, to be presented to the user. In such a way, services emerge from the environment. Our contribution is twofold. The first aims to develop a context-aware distributed engine able to make adequate decisions at runtime about service composition, build and activate adapted composite services and their user-interface. Our engine is an adaptive multi-agent system, where agents learn about composition from experience and user feedback. The second contribution consists of presenting on the fly an emergent service to the user and let her or him modify and/or validate it. Hence, a model of the service must be provided in a user-friendly language depending on her or his profile and skills (e.g., pedestrian, bus driver), using domain specific languages and model-driven engineering. After modification and validation, the model is transformed into both executable re-composition commands and feedback information to supply agents’ learning material.