The focus of evaluation is shifting from the quality of model text output to the ability of agents to complete end-to-end tasks, with evaluation dimensions expanding from a single accuracy metric to multi-dimensional metrics including task completion rate, step efficiency, error recovery capability, and time overhead.