According to CoinMeta, Huang Ruiteng, a senior machine learning engineer at ByteDance and a core contributor, has open-sourced a system designed for long-term tasks. This system enables tasks such as Codex and Claude Code to continue executing towards their original goals even after being interrupted for multiple days. Two real-world task trajectories, spanning 220.7 and 272.9 hours respectively, have been made public. During this period, the system went through multiple rounds of execution, waiting, manual judgment, model switching, and task recovery, yet it was still able to regain its current goal, existing evidence, and next steps. The system separates goals, to-dos, permissions, evidence, and waiting conditions from the model context, making small steps at a time. After verifying the results, it writes them back to the state. Even if the session is changed, the model is switched, or the program is restarted, the task can continue from the latest progress. The component developed by Huang Ruiteng is responsible for saving and retrieving memories, data, and skills related to Agent, while LoopX manages task progress and determines requirements. It is currently being used for automatic code bug fixing, AutoML experiments, and long-term research.