Microsoft has introduced ECHO, a new training method for command-line interface (CLI) agents that enhances learning by predicting terminal outputs. Traditionally, CLI agents were trained to decide the next command based on context, with terminal outputs like errors and logs serving only as background information. ECHO changes this by requiring agents to predict these outputs, thus extracting training value even from failed attempts. Experimental results show significant improvements in training efficiency and performance. On TerminalBench-2.0, Qwen3-8B's pass@1 rate increased from 2.70% to 5.17%, while Qwen3-14B's rate rose from 5.17% to 10.79%. ECHO also reduced training steps by 2.3 times in some experiments and decreased reliance on expert data. However, its effectiveness varies with task complexity, showing notable gains in tasks with clear feedback but a decrease in performance on more complex tasks like TBLite.