One of the most stubborn problems in artificial intelligence is teaching an agent to carry out tasks that stretch far into the future, where rewards arrive rarely and the path to success is measured in hundreds of small decisions. A robot asked to stack blocks into a pyramid or open a microwave must chain together dozens of primitive motions, and a learning algorithm that treats every one of those motions as an independent choice quickly drowns in the sheer scale of the search. Researchers at Shanghai Jiao Tong University and collaborating institutions have now unveiled a framework that tackles this problem head-on by letting an agent learn not only which skills to deploy, but also how long each skill should last, adapting that timing on the fly as circumstances change.