LoopX — Local Control Plane for Long-Running AI Agent Work

• URL: https://github.com/huangruiteng/loopx
• Date Saved: 2026-08-03
• Source: GitHub
• Tags: ai-engineering, dev-tools
• Repo: https://github.com/huangruiteng/loopx

Summary

LoopX is a lightweight state kernel / control plane that sits above AI agent runtimes (Codex, Claude Code, Cursor, etc.) to manage cross-session continuity for long-running agent work. 831 stars, Python, MIT license.

Problem it solves:

A single agent can complete a task in one session. But multi-day/multi-week work is harder: goals change, human decisions appear, evidence expires, peer agents need handoffs, and schedulers waste tokens when no valid state transition exists.

Core state model:

• Goals — what we’re trying to achieve
• Gates — explicit human decision points (not vague “waiting for owner”)
• Todos — ordered, with ownership, claims, and leases
• Evidence — what happened, verified writebacks
• Quota — should the agent run again? budget awareness
• Handoffs — typed continuation between peer agents

Key design choices:

• Agent-agnostic: Codex App, Codex CLI, Claude Code, Cursor, shell agents all supported
• Local-first, no cloud dependency
• No runtime deps beyond Python 3.11+ stdlib
• Peer agents (no leader agent needed) — todo claims, leases, capability gates decide who executes next
• “Executable kanban” mental model — cards have stable identity, permissions, evidence, continuation

Real evidence:

• OpenViking Issue-Fix: 200+ hours natural time, continuous PR delivery
• Auto ML Experiment: 200+ hours, hypothesis/evidence/promote/stop gates
• Auto Research: proposer/executor/evaluator agents in parallel

Relevance:

Addresses the orchestration layer gap for multi-agent teams. Complements the “team knowledge layer for AI” concept — LoopX handles state persistence and handoffs, while a knowledge layer would handle shared context and learning.