Fault handling in Multi-Agent Systems (MAS) is not much addressed in current research. Normally, it is considered difficult to address in detail and often well covered by traditional methods, relying on the underlying communication and operating system. In this paper it is shown that this is not necessarily true, at least not with the assumptions on applications we have made. These assumptions are a massive distribution of computing components, a heterogeneous underlying infrastructure (in terms of hardware, software and communication methods), an emerging configuration, possibly different parties in control of sub-systems, and real-time demands in parts of the system. The key problem is that while a MAS is modular and therefore should be a good platform for building fault tolerant systems, it is also non-deterministic, making it difficult to guarantee a specific behaviour, especially in fault situations. Our proposal is to introduce sentinels to guard certain functionality and to protect from undesired states. The sentinels form a control structure to the MAS, and through the semantic addressing scheme they can monitor communication, build models of other agents, and intervene according to given guidelines. As sentinels are agents themselves, they interact with other agents through agent communication. The sentinel approach allows system developers to first implement the functionality (by programming the agents) and then add on a control system (the sentinels). The control system can be modified on the fly with no or minimal disturbance to the rest of the system. The present work is conducted in cooperation with Sydkraft, a major Swedish power distribution company. Examples are taken from that venture, and it is shown how problems can be solved by programming DA-SoC agents, developed here.