LoopX Open-Sourced: Agents Run 272 Hours Continuously Without Deviation

ByteDance engineer open-sources LoopX control system enabling AI agents to maintain objectives over multi-day tasks, demonstrating stability in 272-hour trajectories without context loss.

Woofun AI reports that Huang Ruiteng, a senior machine learning engineer at ByteDance and core contributor to OpenViking, has open-sourced LoopX. This control system enables long-range agents, such as Codex and Claude Code, to persist in their original objectives across multi-day tasks and interruptions. LoopX abstracts objectives, permissions, and evidence from model context, allowing agents to validate small steps before updating state. This architecture ensures continuity across session switches, model changes, or program restarts.

Two real mission trajectories lasting 220.7 and 272.9 hours have been publicly released, showcasing the system's ability to handle execution, waiting, human judgment, and task recovery. While OpenViking manages memory and skill retrieval, LoopX functions as a project manager tracking task progress and next steps. The tool is currently applied to automatic code issue fixing, AutoML experiments, and long-term research.