Intelligent agents in open environments inevitably face the Qualification Problem: The executability of an action can never be predicted with absolute certainty; unexpected circumstances, albeit unlikely, may at any time prevent the successful performance of an action. Reasoning agents in real-world environments rely on a solution to the Qualification Problem in order to make useful predictions but also to explain and recover from unexpected action failures. Yet the main theoretical result known today in this context is a negative one: While a solution to the Qualification Problem requires to assume away by default abnormal qualifications of actions, straightforward minimization of abnormality falls prey to the production of anomalous models. We present an approach to the Qualification Problem which resolves this anomaly. Anomalous models are shown to arise from ignoring causality, and they are avoided by appealing to just this concept. Our theory builds on the established predicate logic formalism of the Fluent Calculus as a solution to the Frame Problem and to the Ramification Problem in reasoning about actions. The monotonic Fluent Calculus is enhanced by a default theory in order to obtain the nonmonotonic approach called for by the Qualification Problem. The approach has been implemented in an action programming language based on the Fluent Calculus and successfully applied to the high-level control of robots.