This site is built by a loop
looprepo catalogs agent loops — and it is one. The pages you're reading are scraped, graded, published, distributed, and audited by the same kind of loops in the directory, running under the same rules we grade everyone else on.
How the loop runs
Every day a scheduled run goes scrape → ingest → distribute → health. Candidates never skip the default-fail evaluator gate. Longer-horizon work — guides, audits, product changes — runs as agent sessions that read a progress log, do one bounded slice, verify it, and append what they did for the next session. Exit conditions, iteration budgets, a verification step, and a maker-that-isn't-the-grader: exactly what the Safety Score asks of your loops.
The loop, piece by piece
This board is live-edited by the loop itself. Some pieces are still in flight — this page doesn't get to call itself finished until they aren't, because declaring victory before verification is the first failure mode our guides warn about.
| Piece | What it does | Status |
|---|---|---|
| Discovery & scrape | A daily cron sweeps X, GitHub, and the open web for new loop prompts and harness patterns, and writes candidates to a staging table — never straight to the directory. | running |
| Ingest & the evaluator gate | Every candidate passes through a default-fail safety evaluator before it can publish. The gate that grades your loops is the same gate that decides what this site is allowed to add to itself. | running |
| Safety grading (A–F) | One deterministic scorer is the single source of truth for the grader, ingest, the MCP tools, and backfills — so a grade means the same thing everywhere. | running |
| Guides pipeline | Guide drafts are written by agent sessions, gap-checked against outside curricula, then staged as markdown and published by script. The loop-engineering series grew from 12 to 16 guides this way. | in flight |
| Distribution (Postiz) | Social scheduling via Postiz is being wired in so the distribution loop can queue posts on its own. Draft-only until it proves itself — no loop here gets send rights on day one. | in flight |
| Email alerts (outbox → send) | Mission Control run alerts are written to an outbox today. The actual send flips on once credentials land — the loop drafts, a human arms it. | in flight |
| Weekly self-audit | A scheduled session re-reads the site like a stranger would: broken filters, stale tool lists, localhost canonicals. Findings go back into the queue as work items. | running |
| Health checks | The daily cron ends with a health pass over the deploy — the loop checks its own work before anyone else has to. | running |
This page's own harness
Because a loop artifact should carry its guardrails in the open:
- Exit condition: every row above reads running. Then — and only then — this page is done.
- Iteration cadence: updated whenever a piece changes state. Last iteration: 2026-07-12.
- Verification: a status only flips to running after the piece has been observed working in production, not when the code merges.
- Scope guard: the loop can edit this manifest; it cannot weaken the evaluator gate, touch billing, or grant itself send rights. Those stay human-armed.
Want a loop with the same shape for your own project? Start with the harness builder or browse graded loops.