# LoopX: A Local Control Plane for Long-Running Agent Work

> LoopX keeps goals, evidence, ownership, quotas, and handoffs stable across long-running agent sessions without replacing the agent runtime.

- Author: Kai Wang (AI Kai)
- Published: 2026-08-23
- Updated: 2026-08-23
- Topic: Agent Workflows
- Tags: loopx, multi-agent, agent-ops
- Canonical: https://hqman.me/blog/loopx-control-plane/

## The problem

A single agent can finish a task in one session. The real trouble starts when work spans days, multiple sessions, and multiple agents. Objectives shift. Evidence expires. Tasks go unclaimed. A scheduler can continue consuming quota after the workflow has no valid transition left.

Chat memory and timers don't govern any of this.

## What LoopX does

LoopX is an open-source local control plane that sits outside your agent runtime (Codex, Claude Code, Cursor, Pi, or your own). It doesn't replace the agent. It keeps the long-term state stable between turns.

The agent handles one bounded slice of work per turn. LoopX holds everything else: the objective, gates, todos, scope, evidence, quota, and handoff state. Its gates and quota checks indicate whether the next tick should continue, wait, or ask for human judgment.

```
Objective / issue / project
   |
   v
LoopX state: objective + gate + todo + scope + evidence + quota
   |
   |-- Needs human judgment? -- yes --> Ask a concrete question and wait
   |
   |-- Has a safe path? -----------> Execute one bounded agent slice
   |
   v
Agent executes one turn (Codex / Claude Code / Cursor / shell)
   |
   v
Write back evidence + handoff + next todo --> quota decides next tick
```

## Multi-agent coordination

Multiple agents can claim different tasks through ownership, lease, and handoff. These mechanisms can reduce conflicting edits and make stalled work visible. Think of it as an "executable kanban board" for long-running agent work: recoverable, auditable, and ready for the next agent to pick up.

## Real traces

The public README documents an issue-fix trajectory and an owner-run ML experiment that each span more than 200 hours of elapsed time. The claim describes wall-clock project time, not continuous model execution or unattended production autonomy.

## Why it matters

Keeping a model running is only one part of long-running agent work. The objective, evidence, ownership, and next step also need to survive between turns.

LoopX is still early. Its README describes the v0.4.x line as an early but usable local control plane, not a full agent platform or autonomous production controller.

![LoopX control plane](https://hqman.me/media/blog/loopx-control-plane/loopx-control-plane.webp)

[View LoopX on GitHub](https://github.com/huangruiteng/loopx)