Open multi-agent systems (MAS) are decentralized and highly distributed systems that consist of a large number of loosely coupled autonomous agents. Diagnosing exceptions in such systems is a complex task due to the distributed nature of their data and their control. This complexity is exacerbated in open environments where independently developed autonomous agents interact with each other in order to achieve their goals. Inevitably, exceptions will occur in such MAS and these exceptions can arise at one of three levels, namely environmental, knowledge and social levels. In this paper we propose a novel exception diagnosis system that is able to analyse and detect exceptions effectively. The proposed architecture consists of specialised exception diagnosis agents called sentinel agents. The sentinel agents are equipped with knowledge of observable abnormal situations, and their underlying causes.