A developer using the GitHub handle huangruiteng has released LoopX, an open-source utility that keeps a persistent record of what an AI agent should be doing across sessions stretching over days or weeks. The tool does not run models or write code itself. It manages the surrounding paperwork: objectives, pending tasks, evidence of what happened, and notes for whichever agent picks up the work next.
Multi-agent systems built on Codex, Claude Code, or Cursor rarely fail because the underlying model reasons poorly. They fail because nobody, human or agent, can say with confidence what the current objective is, who owns the next step, or whether a change made three sessions ago was ever verified. That is a state management problem, not an intelligence problem, and it is the unglamorous work that determines whether a long-running agent team finishes anything at all.
LoopX stores that state locally in a compact layer the project calls a kernel: an active goal, a set of gates that require a human decision, an ordered list of todos, a log of evidence, and a quota that governs how often an agent may act. Before taking a turn, an agent checks whether the loop should run at all, claims a todo, does one bounded piece of work, then writes back evidence and a handoff note for whichever turn comes next. The project frames the resulting workflow as closer to a Kanban board built for agents than a chat transcript, with claim, gate, and writeback acting as the moves that keep the board honest.
Agents in this model act as peers rather than a hierarchy: ownership passes through claims and leases instead of a single controlling agent deciding what happens next. The project’s own documentation points to two worked examples, an issue-fix contribution to a related open-source project called OpenViking and a machine-learning experiment loop, both of which it says ran across more than 200 hours of elapsed project time. Elapsed time here spans calendar days of intermittent, bounded turns rather than 200 unbroken hours of a model running, a distinction the documentation itself takes care to draw.
Installing LoopX is a single curl command that pulls a Python package with no external dependencies, and it connects to Codex, Claude Code, Cursor, and custom runners through separate adapters. LoopX’s own status page describes the project as version 0.4.x: early but usable, and explicitly not a full agent platform or an autonomous production controller. The documentation is candid that the tool issues no credentials, signs off on nothing in production, and posts nothing under a user’s name, and it offers no independent adoption numbers beyond a self-hosted star count graph, a reasonable caution for a project with no vendor backing or outside audit behind it.
Teams running Codex or Claude Code across multi-day engineering or research tasks tend to lose more time to lost context and unverifiable handoffs than to weak model output, which makes a free, dependency-light state layer worth a trial run before building the same bookkeeping in-house. The open question is whether a single-maintainer project can keep pace with the release cadence of the agent runtimes it wraps.
Based on the LoopX repository maintained by huangruiteng on GitHub, viewed August 7, 2026.
