The Missing Control Plane for Long-Running Agent Loops

Single-session agents are fine for one-shot tasks. The trouble starts when you hand Claude Code something that takes days — audit this whole repo, fix 40 issues across a sprint, monitor a moving target. By the next terminal session, it starts from scratch: re-reads context it already processed, loses track of which tasks are done, and has no idea what questions it was waiting on you to answer. LoopX is the layer that fixes this.

What It Is

LoopX is a lightweight, local-first state kernel for long-running agent work. It sits above your existing agent harness — Claude Code, Codex App, Codex CLI, or Cursor — and maintains durable state across sessions: active goals, an ordered todo list, completed items, human decision gates, evidence trails, and execution quotas. The core idea is what its author calls *loop engineering*: structured, resumable agent runs where human judgment gates the decisions that actually need it, and everything else keeps moving.

When It Shines

The gate mechanism is the part worth paying attention to. Instead of the agent going silent on an ambiguous decision — or worse, hallucinating through it — LoopX surfaces a concrete question and pauses: *Found two approaches here — A or B?* You answer, it continues. The README documents three real use cases: a 200+ hour open-source contribution arc, a 200+ hour ML experiment spanning hypothesis validation and gate decisions, and independent user reports of a four-day unattended run producing seven merged PRs. Practically, it is most useful for: large refactors you cannot finish in one session, recurring monitoring or triage jobs, multi-step research with branching evidence, and any workflow where you want steady agent progress without daily hand-holding.

curl -fsSL https://huangruiteng.github.io/loopx/install.sh | bash
export PATH="$HOME/.local/bin:$PATH"
loopx doctor

One Real Caveat

LoopX has opinions, and they are load-bearing. The loop engineering model — gates, evidence, quota mechanics, presets — is not a thin wrapper you can ignore; it is the point. Expect a session or two learning the mental model before it clicks, and the CLI surface is rough enough that loopx doctor frequently surfaces environment surprises on first run (Python 3.11+ required, PATH setup catches people). If you want something that works on day one with zero onboarding, look elsewhere. If you are building real sustained agent workflows and are willing to invest a couple of hours in the plumbing, the payoff is genuine.

Where to Find It

LoopX is built by huangruiteng and lives at github.com/huangruiteng/loopx. Apache 2.0 license, 4.8k stars, active Discord. The quote at the top of the README sums up the design philosophy well: *"Keep the loop moving. Keep the judgment human."*