A library for agents that know when to stop

What it does Loop Library is a public catalog of iterative AI-agent workflows—called “loops”—plus an optional skill that helps agents find, audit, adapt, or design them. Each loop replaces a one-shot prompt with a bounded cycle: attempt, verify, learn, and either repeat or stop. The website hosts the catalog, while the installable skill lets agents query it during conversation to pick or tailor a loop for a specific task.

The interesting bit The project treats the exit strategy as part of the prompt itself. Each loop is essentially a unit test for agent behavior: it defines the assertion first, then the repetition logic, which keeps agents from confusing motion with progress.

Key highlights

• Public catalog of bounded, repeatable workflows for engineering, evaluation, operations, content, and design
• Each loop specifies what to attempt, how to verify it, what to learn, and when to stop or escalate
• Optional agent skill integrates with Codex, Cursor, and Claude Code to find, audit, adapt, or design loops conversationally
• Agents can consume the catalog directly via JSON, plain text, or dedicated agent guides without installing the skill
• New loops can be published without a site redeploy or repository commit

Verdict Give this a spin if your agents currently treat “keep improving this” as a license to iterate forever. If you only fire off single-turn prompts, you probably do not need a library of playbooks.