ByteDance engineer open-sources LoopX for long-running AI agents

According to Beating Monitor, Huang Ruiteng, Senior Machine Learning Engineer for AML at ByteDance and core contributor to OpenViking, has open-sourced LoopX—a control system designed for long-duration agents, enabling agents such as Codex and Claude Code to continue from their original objectives after tasks span multiple days and multiple interruptions. LoopX has publicly shared two real task trajectories lasting 220.7 and 272.9 hours, respectively. During these periods, involving multiple rounds of execution, waiting, human judgment, model switching, and task recovery, the agent successfully regained its current objective, existing evidence, and next steps. LoopX externalizes goals, to-dos, permissions, evidence, and waiting conditions from the model’s context. The agent performs only one small step at a time, verifies the result, then writes the state back. Whether switching sessions, models, or restarting the program, it can resume from the latest progress. Huang Ruiteng’s OpenViking handles the storage and retrieval of an agent’s memory, data, and skills; LoopX manages task progress, next steps, and when human judgment is required. The former acts like long-term memory, while the latter functions more like a project manager and executable board. Currently, LoopX is being used for automated code bug fixes, AutoML experiments, and long-term research.