Loop Library and Loopy: Feedback-Aware Agent Workflows for Claude Code, Cursor, and Codex

The Problem with Open-Ended Agent Instructions

Most agent prompts tell an AI to accomplish something once. The agent acts, produces output, and stops. That model works for simple, well-defined tasks but breaks down when the first attempt will not be the final answer. Fixing a recurring production error, improving test coverage over multiple passes, or keeping documentation in sync with evolving code all require the agent to look at what it just did and decide what to do next.

Loop Library addresses this gap. It defines a loop as a prompt that also tells the agent how to judge its own output, what to do with what it learned from the last pass, and when to stop or hand control back to a person. Each published loop answers four questions: what the agent is trying to accomplish, how it will know whether the latest attempt worked, what it should do with what it learned, and when it should finish or ask for help.

The repository splits into two distinct parts. The Loop Library website is a public catalog that agents and people can browse or query without installing anything. The Loopy skill is an optional companion that integrates the catalog directly into an agent's available tools on supported platforms.

What Makes a Loop Different from a Prompt

A standard prompt asks for a result. A loop from this library builds feedback into the instruction itself. The README illustrates the difference with a concrete example. A one-shot instruction might say, "Make this website faster." A loop version says: "Find the slowest page, make one focused improvement, and measure it again. Keep the change only if it helps. Repeat until every page meets the target or another pass stops producing a meaningful improvement."

That structure changes how the agent behaves. Instead of making a change and stopping, it compares results across passes, keeps improvements, and terminates when it succeeds or when further passes yield no measurable gain. The README is explicit that loops are not permission for an agent to run forever. The best ones are deliberately bounded: they include a real check, a clear stopping point, and a moment to hand control back to a person when judgment or approval is needed.

This design makes the same loop reusable across different projects or team members without rebuilding the feedback logic from scratch each time.

Installing Loopy and Connecting It to Your Agent

Loopy requires Node.js and npx. Install it globally for Claude Code with this single command:

npx skills add Forward-Future/loopy --skill loopy --agent claude-code -g -y

To install for all three supported platforms at once:

npx skills add Forward-Future/loopy \
  --skill loopy \
  --agent codex \
  --agent cursor \
  --agent claude-code \
  -g -y

The `-g` flag makes Loopy available across all projects on your machine. Omit it to restrict the install to the current project. The `-y` flag accepts the install prompts without interaction; leave it off to review the choices interactively. If you use an agent not listed above, run the interactive installer:

npx skills add Forward-Future/loopy --skill loopy -g

This detects the agents present on your machine and lets you choose. After installation, restart the agent if it was already running. The previous `loop-library` skill name still works as a compatibility alias for existing installs, but all new installs should use `loopy`.

In Claude Code, invoke Loopy with `/loopy` followed by your request. In Cursor, type `/` in Agent chat, search for `loopy`, and select it. In Codex, type `/skills`, choose Loopy, then enter your request, or mention it directly with `$loopy`. The README gives this Codex example:

$loopy Analyze this codebase and my coding threads for repeated work, then turn the strongest candidate into a reliable loop.

You do not need to know loop terminology before starting. Describe what you want to accomplish in plain language and Loopy selects the appropriate path.

Nine Paths and What Loopy Will Not Do Without Approval

Loopy takes nine distinct paths depending on what you ask. It can discover repeated work in a codebase or coding threads and turn the strongest candidate into a loop. It can find a published loop that fits a task, audit an existing loop for weak checks or unsafe actions, adapt a loop to your tools and constraints, or craft a new one through a short plain-language conversation. It can also run a loop in bounded passes and return a receipt listing the actions taken, the evidence gathered, the outcome, and the stopping reason. After a run, it can debrief the results and suggest the smallest evidence-backed improvement. Finally, it can save a loop to the project's `LOOPS.md` file for reuse in later sessions, or prepare a draft for publication in the catalog.

The README is direct about the boundaries. Loopy does not quietly start schedules, change production systems, publish content, or send messages on behalf of the user. Those actions still require the same permissions and approvals that apply to the agent platform itself. This is a meaningful constraint for teams evaluating whether to give an agent broader access: Loopy acts as a guided interface to the catalog and as a structured execution wrapper, not as an autonomous operator.

Repository Layout and the Catalog Without Loopy

The repository has three main code paths. The Loop Library website code lives under `loop-library/`, split between `loop-library/site/` for the front-end shell and `loop-library/worker/` for the database and rendering logic. The Loopy skill source lives under `skills/loopy/`.

Agents that do not have Loopy installed can still use the catalog. The README documents four direct access points: an agent guide, a plain-text instructions file at `llms.txt`, a JSON catalog at `catalog.json`, and a plain-text catalog at `catalog.txt`, all hosted under the signals.forwardfuture.com/loop-library path. Installing Loopy adds the guided workflow but does not install or host the website. This separation means that the catalog is available to any agent that can make an HTTP request, regardless of platform.

Comparison with Programmatic Agent Frameworks

The closest programmatic alternative is a framework like LangChain, which is a Python-based library for building applications that chain LLM calls together. The difference in approach is fundamental. LangChain encodes the loop logic in source code that a developer writes and maintains. Loop Library stores the same logic as plain-text prompts that any supported agent can read and follow without writing or deploying code. This makes Loop Library accessible to people who do not write Python but gives it less flexibility for cases that require complex conditional branching, stateful session management, or dynamic tool composition at runtime. LangChain also runs as a Python package inside your infrastructure; Loop Library fetches the catalog from an external website, which introduces a dependency on that service being available.

Maintenance, Release History, and License

The repository has no GitHub releases. The last push was on 2026-09-11. The project is MIT-licensed. The previous skill name `loop-library` remains available as a compatibility alias for users who installed it under that name, but the README states that `loopy` is the name to use for all new installations and explicit invocations.

Editorial conclusion

Loop Library suits developers and teams that run the same agentic task more than once and want a shared, audited playbook rather than a fresh ad-hoc prompt each time. It is not the right fit if your agent work is mostly one-shot and you have no need to compare successive results across passes. Before adopting it, confirm your agent platform is one of the three that Loopy currently supports: Codex, Cursor, or Claude Code. Agents on other platforms can still access the catalog through the public llms.txt and catalog.json endpoints without installing the skill.

Frequently asked questions

Official sources

1. Forward-Future/loopy on GitHub
2. Issues
3. License: MIT
4. Project website
5. README