Huang Ruiteng, a senior machine learning engineer on ByteDance's Applied Machine Learning (AML) team and a core contributor to OpenViking, has open-sourced LoopX. It's a control system for long-horizon agents that keeps Codex, Claude Code, and their peers moving toward the original goal even when a task spans multiple days and suffers repeated interruptions.
LoopX has already published two real task trajectories spanning 220.7 and 272.9 hours. During those runs, agents went through repeated execution rounds, waiting periods, human judgment calls, model switches, and task recovery — and still managed to find the current objective, the evidence on hand, and the next step.
LoopX moves goals, to-dos, permissions, evidence, and wait conditions outside the model's context. The agent takes one small step, verifies the result, then writes it back to the state. Swap sessions, swap models, or restart the program completely — it can always pick up from the latest progress.
OpenViking, which Huang also built, handles storing and retrieving an agent's memories, materials, and skills. LoopX manages how far along a task is, what should happen next, and when a human needs to jump in. The former is like long-term memory; the latter is more like a project manager with an executable kanban board. LoopX is already being used for automatic code fixes, AutoML experiments, and long-term research.
https://twitter.com/huangruiteng/status/2083904257494024425