Reinforcement learning (RL) has become a key approach for enhancing large language model (LLM) agents, endowing them with stronger reasoning, planning, and tool‑use in interactive settings. Existing agent‑RL methods, however, suffer from an on‑policy exploration bottleneck: when tasks lie beyond the model's current capability, successful trajectories are extremely rare and training often collapses.\ \ We introduce Rollout‑Time Intervention (RTI), which leverages reward‑side evidence (e.g., evaluation rubrics) during rollouts to transform erroneous trajectories into successful ones, thus supplying effective learning signals for tasks the policy rarely solves on its own.\ \ Although intervention trajectories are inherently off‑policy, filtering out those with excessive distributional shift allows them to guide policy optimization effectively.\ \ Extensive experiments on five benchmarks show that RTI consistently outperforms RL baselines, enabling the post‑trained policy to approach the performance of substantially stronger contemporary LLMs and, more importantly, to transcend its original capability boundary. The implementation is publicly available at https://github.com/Alibaba-NLP/qqr.\ \ Review