In the screenshot-recognition approach, the agent periodically takes screenshots and uses a vision model to locate targets such as buttons and text boxes, then simulates mouse and keyboard events; it is more adaptable, but accuracy depends on model understanding, and resolution changes, animation transitions and pop-up occlusion may cause misjudgments, with every operation accompanied by visual reasoning, resulting in higher latency and cost.