Code Agents performing complex terminal tasks often get stuck in ineffective operations after making errors and struggle to learn from those failures. To tackle this problem, researchers propose ReTerm, a framework that explores a recursive self-improvement path driven by the agents' own mistakes. The method first builds a recovery demonstration dataset at the ten-thousand scale for cold start, enabling the model to learn recovery operations from failure situations it generates itself. Then, through ReTerm-GRPO, the agent autonomously attempts to repair errors and converts repair success or failure into reinforcement learning signals to update its parameters, forming a closed loop of "producing failures — attempting recovery — iterative optimization." Experiments show that the 9B-parameter ReTerm model outperforms the base model by more than 13 percentage points on multiple evaluations, including terminal operations and software engineering. These gains transfer to different runtime frameworks and generalize to areas such as mathematical reasoning and table operations. Recovery rates for multiple error types exceed those of both the base model and conventional reinforcement learning schemes, confirming that learning to recover from one's own failures is an effective path to improving terminal agent capabilities and offering a practical experimental entry point for failure-driven recursive self-improvement.