This paper presents an agent-oriented approach to build a decision support system aimed at helping emergency managers to detect and to manage risks. We stress the flexibility and the adaptivity characteristics that are crucial to build a robust and efficient system, able to resolve complex problems. The system should be independent as much as possible from the subject of study. Thereby, an original approach based on a mechanism of perception, representation, characterisation and assessment is proposed. The work described here is applied on the RoboCupRescue application. Experimentations and results are provided.