The most consequential line in the LoopX README is a disclaimer: "LoopX is not an autonomous production controller." Dangerous permissions, publishing, production writes, and final ownership stay with the human. For a project whose whole subject is agents that run for days, that is a real constraint, and the rest of the repo is built to honor it.
The problem is the gap between turns
LoopX starts from an observation most agent tooling skirts. A single session can finish a task. Long-running work breaks somewhere else: objectives shift, an owner decision shows up mid-stream, evidence goes stale, one agent hands work to a peer, and a scheduler keeps burning turns after there is nothing useful left to transition to. The README's phrasing is blunt about the usual answer. Chat memory and a timer are not enough to govern that.
So LoopX does not try to be the thing that executes. It calls itself a lightweight state kernel and a local-first control plane, running on top of Codex App, Claude Code, Cursor, dsh, or a harness you wrote yourself. The harness executes one bounded turn. LoopX holds the objective, the gates, the todos, the scope, the evidence, and the quota, then decides what happens next. Stated plainly in the docs: not another agent framework, not a provider-specific orchestration runtime.
The control flow diagram is short enough to read in one pass. State comes in. If human judgment is needed, the loop asks a concrete question and waits. If a safe fallback exists, it runs one bounded agent slice. The runtime takes its turn, writes evidence, a handoff, and the next todo, and quota decides whether there is another tick.
That waiting step is the part worth stealing even if you never install this. Most loops fail open. This one is designed to stop and ask.
An agent-native Kanban, with the board demoted
The mental model the docs offer is a Kanban board for long-running work, but with a twist: cards carry identity, authority, evidence, and continuation, and moves are validated operators such as claim, gate, monitor, and writeback. The board is a projection. LoopX state stays the source of truth.
The same rule governs the workspace UI. loopx dashboard launches a browser or PWA surface where you can see what needs you, what is running, what is being watched, and what is scheduled or stopped, then continue across Codex, Claude Code, or direct-model sessions without losing Goal state. Protected changes go through typed preview, explicit confirmation, and receipts. The README is careful to note that LoopX state, not the browser, remains authoritative. The experimental Tauri desktop shell gets the same treatment: it reuses the loopback status and Chat services rather than becoming a second state authority, and closing the window stops only what the shell itself started.
Registered agents are peers. Claims, leases, task boundaries, capabilities, and typed continuation decide who acts next. No durable leader identity required.
Evidence with the caveats attached
The showcase section is where the project's temperament shows clearest. Two arcs, a public OpenViking contribution sequence and a redacted owner-run Auto ML experiment, each span more than 200 hours of elapsed loop lifetime. The README immediately defines the term: elapsed lifetime is wall-clock project time, not 200 hours of continuous model execution, and not a claim of unattended production autonomy. The ML graph is labeled an owner-run showcase, not a reproduction, not a production result, not an employer endorsement.
The third case, Auto Research, is the reproducible one. Proposer, executor, and evaluator/promoter agents iterate in parallel over a built-in exact-KNN demo whose task definition, editable and protected files, deterministic CPU evaluator, and dev/held-out commands all ship in the repository. Even there the README calls it a demo result.
Independent user reports carry the same labeling: a C++ accuracy run past the thirteen-hour mark, a four-day unattended run, seven merged PRs on a public engine refactor with the reported token scale marked as a user claim rather than a measurement.
Fitting this to your own work is a short test. If your agents finish inside one session, skip it. If you run multi-day objectives, issue and PR loops that must preserve scope and review state, gates for safety or publication or private data, or peer agents that need leases and handoffs, this is the layer that has been missing.
GitHub: https://github.com/huangruiteng/loopx
Curated by Agent Palisade — practical AI for small and mid-sized businesses.
Top comments (0)