Researchers introduce ComputerSD, an online self-distillation method for computer-use agents (CUAs) that learns from real-time feedback from GUI transitions. ComputerSD outperforms existing methods by 1.9-4.1 percentage points on two benchmark backbones.