Forward-Future/loopy — a catalog and skill for bounded, reusable agent loops

TL;DR

Forward-Future/loopy is two things in one MIT-licensed repo (2.2k★, JavaScript): Loop Library, a public, machine-readable catalog of practical agent loops, and Loopy, an installable cross-agent skill (Codex / Cursor / Claude Code via npx skills add) that discovers, finds, audits, adapts, crafts, runs, debriefs, and publishes them. Its load-bearing idea is the loop: a one-shot prompt turned into a bounded, repeatable workflow that answers four questions — goal, how-to-verify-progress, what-to-do-with-the- learning, when-to-stop-or-ask — anchored by a real acceptance check and a deliberate hand-back to a human. This is Anthropic's evaluator-optimizer / bounded-agent-loop discipline (Agentic Workflow Patterns, Workflows vs Agents), but productized: loops become named, self-describing, publishable artifacts (Use when / Prompt / Verify / Steps / Notes / Related) living in a shareable catalog. Worth caring about on two counts — as an independent corroboration that agent loops must be bounded and checked, and for its genuinely new angle: treating repeatable workflows as reusable library artifacts rather than one-off prompts. The website/worker plumbing (Cloudflare Worker + SQLite Durable Object, append-only revisions, fail-closed OAuth voting) is engine detail — held, not distilled.

What it is & how it is organized

• A GitHub repo, not prose — read at triage altitude. Two separable parts:
  • loop-library/ — the website/catalog (a site/ shell + a Cloudflare worker/ that holds the loop records in a SQLite-backed Durable Object and renders JSON / Markdown / plain-text / llms.txt catalogs from a single database write). The catalog is the library; no install required to browse it, and agents can read it directly.
  • skills/loopy/ (with a loop-library compat alias) — the installable skill that guides an agent through the loop lifecycle against that live catalog.
• The Loopy lifecycle is eight paths: Discover (mine recurring work → candidate loop), Find (recommend up to three published loops), Loop Doctor (audit + repair material problems only), Adapt (fit to local tools/limits), Craft (interview → new loop), Run (bounded passes + acceptance check + evidence receipt), Debrief (one minimal justified change), Publish (overlap-check + explicit-approval gate).
• Safety posture is explicit: Loopy does not start schedules, change production, publish, or send messages without normal permissions and approvals; a finite run boundary is required before acting.

Concepts introduced

• Agent Loop — the bounded, feedback-driven workflow itself: four-question anatomy, a real acceptance check, explicit stop conditions, and hand-back to a human. The evaluator-optimizer pattern promoted to a first-class design object.
• Reusable Workflow Library — agent workflows as durable, cataloged, publishable artifacts with a fixed self-describing anatomy, exposed in machine-readable form for agents to consume directly.

Held, not dropped

Themes the capture touches that do not warrant their own concept page yet — spin out on demand:

• Cross-agent skill packaging / distribution — npx skills add … --agent {codex,cursor, claude-code}, skills as installable guides portable across agents. (A distribution mechanic, adjacent to but distinct from the loop concept.)
• Machine-readable agent catalogs — llms.txt, catalog.json, an agents guide as the interface an agent reads instead of a human UI. (Touches Agent-Computer Interface (ACI); held as a general pattern.)
• Catalog publishing infrastructure — Cloudflare Worker + SQLite Durable Object, append-only revision per publish, draft/archive states, enforced bootstrap digest. (Engine plumbing, not LLMOps pedagogy.)
• Fail-closed staged rollout & OAuth security — HMAC nonce-bound state, no-store callback bridge, VOTING_UI_ENABLED flag flip after a smoke test. (Web-app security, out of scope.)

Key claims

• A loop is a playbook with feedback built in — repeatable, not one-shot. principle → Agent Loop.
• A good loop answers four questions: goal, verify-progress, use-the-learning, stop-or- ask. best practice — context: designing a bounded iterative workflow → Agent Loop.
• Loops are deliberately bounded — a real check, a finite run boundary before acting, and a hand-back to a human at approval/judgment boundaries; "not permission to run forever." principle → Agent Loop, corroborating Workflows vs Agents.
• Require ≥2 distinct occurrences before calling work "repeated"; a code pattern without run history is a potential loop, not proven recurrence. best practice — context: promoting recurring work into a loop → Agent Loop.
• Agent workflows are reusable artifacts worth naming, cataloging, and sharing — not disposable prompts. principle → Reusable Workflow Library.
• A shareable workflow needs a self-describing anatomy (use-when / prompt / verify / steps / notes / related) and a machine-readable catalog for agents to consume. best practice — context: portability across people and agents → Reusable Workflow Library.
• A debrief separates loop-design failures from execution / tool / environment / goal failures before recommending one minimal change. observation — echoes Error Analysis's discipline of attributing failure to the right layer.

Why this builds on (and corroborates) the existing graph

This is the dominant stance — builds_on, with a strong secondary corroborates:

• Builds on Agentic Workflow Patterns and Workflows vs Agents: the "loop" is the evaluator-optimizer pattern and the bounded agent loop, but extended into a named, reusable, publishable artifact with a fixed anatomy and lifecycle. That artifact-and-reuse layer (Reusable Workflow Library) attaches to nothing in the prior graph — the genuinely novel contribution.
• Corroborates the bounded-autonomy discipline: Forward Future is an independent source (not Anthropic) arriving at the same load-bearing claims — loops must be bounded, must have a real acceptance check, and must hand control back to a human. Two sources now agree, which is exactly why the claim is recorded as a (principle) rather than one vendor's (best practice). No concept was duplicated to absorb this; the existing pages gained corroborating claims and a backlink instead.

Linked from

• Agent Loop
• Agentic Workflow Patterns
• Finally. Agent Loops Clearly Explained. — loop engineering, decoded for the rest of us
• Forward-Future/loop-library — the monorepo behind Loopy (duplicate of Forward-Future/loopy)
• This Week
• Reusable Workflow Library
• Workflows vs Agents