loopy: a framework for creating and managing iterative feedback loops for AI agents

What it solves

It addresses the limitation of "one-shot" prompting, where AI agents are asked to perform a task once without a mechanism for iterative improvement. By introducing "loops"—playbooks with built-in feedback—it enables agents to learn from results, verify their work, and repeat steps until a specific goal is met or progress stalls, making agentic workflows more reliable and repeatable.

How it works

The project consists of two components: a public Loop Library (a catalog of published loops) and Loopy, an installable skill for AI agents (compatible with Codex, Cursor, and Claude Code).

Loopy allows agents to:

• Discover: Identify recurring patterns in a codebase or chat history to create new loops.
• Find & Adapt: Search the library for existing loops and tailor them to specific project constraints.
• Craft: Use an interview process to build a custom loop from scratch.
• Run & Debrief: Execute loops in bounded passes and analyze the results to suggest minimal improvements.
• Save & Publish: Store loops locally in a LOOPS.md file or submit them to the public library.

Who it’s for

Developers and AI agent users who want to move beyond simple prompts to structured, iterative workflows for tasks like fixing production errors, improving test coverage, or maintaining documentation.

Highlights

• Bounded Execution: Loops include clear stopping points and approval boundaries to prevent agents from running indefinitely.
• Evidence-Based: Requires verification of success and provides "receipts" of actions and outcomes.
• Agent Integration: Directly installable as a skill into popular AI coding agents.
• Project-Local Storage: Supports saving custom loops to a project's root directory for reuse across sessions.