“To improve AI agent capabilities, Ali Ansari argues that training data needs to shift from simple question-answer pairs to long-horizon tasks that simulate complete, end-to-end workflows.”