In this work, we present a new intuitive, end-to-end approach for temporal action detection in untrimmed videos. We introduce our new architecture for Single-Stream Temporal Action Detection (SS-TAD), which effectively integrates joint action detection with its semantic sub-tasks in a single unifying end-to-end framework. We develop a method for training our deep recurrent architecture based on enforcing semantic constraints on intermediate modules that are gradually relaxed as learning progresses. We find that such a dynamic learning scheme enables SS-TAD to achieve higher overall detection performance, with fewer training epochs. By design, our single-pass network is very efficient and can operate at 701 frames per second, while simultaneously outperforming the state-of-the-art methods for temporal action detection on THUMOS'14.