LoopX: An Open, Stateful Control Plane for Long-Horizon Agents

A Control Plane, Not Another Agent Framework

LoopX describes itself in one sentence: an open, provider-neutral, stateful control plane for long-horizon agents. It runs on top of Codex, Claude Code, Cursor and other agent harnesses instead of replacing them, and adds exactly the layer those harnesses lack: objectives, gates, todos, scope, evidence, quota and handoffs that stay durable across turns, restarts and even across different agents. The README's mental model is an agent-native Kanban for long-running work, where cards carry identity, authority, evidence and continuation, and every move is a validated operator such as claim, gate, monitor or writeback.

The motivating argument is concrete. An agent can finish a task inside one session, but multi-day work fails in specific ways: objectives drift, owner decisions appear mid-run, evidence goes stale, work is handed between peer agents, and a scheduler keeps spending quota after no useful transition remains. Chat memory plus a timer cannot govern that. LoopX therefore keeps the durable control state in one compact kernel and lets the harness execute bounded slices of work.

The Kernel's Five Questions

The architecture folds into five questions, each mapped to a product promise on top of any harness: what is the objective (active goal, explicit scope, current authority); what happens next (ordered user and agent todos with ownership, claims and leases); what needs human judgment (concrete user gates instead of a vague "waiting for owner"); what evidence changed (compact run history, validation, blockers, accepted writeback); and may the loop continue (quota, capabilities, safe fallback, scheduler hints, stop conditions). Dangerous permissions, publishing, production writes and final ownership explicitly stay with the human.

The runtime contract is equally explicit. The execution path is Agent to Capability to Provider; the control path returns Provider readback to Capability transition to Kernel. A capability is a stable, provider-neutral contract that produces one bounded, verifiable outcome; a provider calls an external system and returns observations and readback; an extension packages an optional provider with an explicit install, enable, upgrade, disable and rollback lifecycle. Registered agents are peers: claims, leases, task boundaries and typed continuation decide who acts next, with no durable leader identity required.

Quota as the Governor

The most distinctive primitive is quota. loopx quota should-run answers whether a registered agent may act now, and the interaction contract decides whether a turn should deliver, ask, wait, self-repair or stay quiet. The minimal custom-runtime tick is five commands: quota should-run, todo claim, todo update, refresh-state, quota spend-slot. That tiny surface is the whole loop driver a custom runner needs, which is why the host table spans Codex App, Codex CLI, Claude Code, KunlunCode, OpenCode, Pi, ZCode, Antigravity CLI, DeepSeek Harness, Cursor and plain shell runners without per-host forks of the governance logic.

Evidence Over Claims

LoopX is unusually careful about evidence strength. Its flagship cases are a 200+ hour public OpenViking contribution arc (the creator's own PR sequence, with rolling repository context and revision-stamped fix knowledge), a redacted 200+ hour owner-run AutoML experiment graph that keeps hypotheses, matched evidence, invalid lineages and promote/stop gates visible, and a reproducible in-repo exact-KNN auto-research demo where proposer, executor and evaluator/promoter agents iterate in parallel. Independent user reports add a 13-hour C++ accuracy run, a four-day unattended run and seven merged PRs on a public engine refactor; every case page states its reproducibility boundary and explicitly denies continuous-compute or production-autonomy claims.

Getting In

Installation is pip install loopx plus loopx workflow-skills --install and loopx doctor; then loopx connect in a project root, or the guided loopx start-goal when state is missing. Python 3.11+ and Node 22.6+ are required, the latter for the managed, idle-exiting TypeScript Effect core that LoopX starts automatically. loopx dashboard opens the local-first Personal Agent Workspace, where goals, gates, evidence, scheduled watches and Lark-routed inbox messages live; the browser is a projection and never the state authority. Shipped domain capabilities include issue fixing, change-quality qualification, integration-branch reconciliation, exploration, decision context and periodic reports. Apache-2.0, roughly 5.7k stars at ingest time.