Existing work in AU detection typically formulate the problem as classification between frames or segments of positive and negative examples, and emphasizes the use of different features or classifiers. This paper proposes a novel AU event detection method, Cascade of Tasks (CoT), which combines the use of different tasks (ie, frame-level detection, segment-level detection and transition detection). We train CoT sequentially embracing diversity to ensure robustness and generalization to unseen data. Unlike conventional frame-based metrics that evaluate frames independently, we propose a new event-based metric to evaluate detection performance at the event-level. We show how the CoT method consistently outperforms state-of-the-art approaches in both frame-based and event-based metrics, across four datasets that differ in complexity.