wtf is loop-engineer?

This project is a skill and subagent for Claude Code, Anthropic's coding assistant tool, that helps a developer set up what the README calls agentic loops: automated cycles where an AI agent runs repeatedly against a task instead of being prompted once by a human. Rather than inventing new loop tools, this project scaffolds and adds safety limits around loop features that already exist in Claude Code, such as running it headlessly, its built in loop and goal commands, tasks, and scheduled cloud routines. The README describes five loop patterns: a simple loop that checks one pass or fail condition, an evaluator-optimizer pattern where a separate step judges the output of a generating step, a meta loop that improves the prompt another loop uses, an orchestrator that fans work out to multiple workers when the subtasks cannot be predicted ahead of time, and a scheduled loop that runs on a timer. Each pattern has a runnable shell script template. The README states these patterns can be nested, such as an orchestrator that spawns evaluator-optimizer loops inside it. The central rule stated in the README is that a loop must have a verifiable, binary exit condition, since a vague goal like improving user experience has no clear pass or fail state and can run indefinitely while accumulating cost. It cites an example of a four-agent loop running eleven days and costing about forty seven thousand dollars because it lacked this check. Every scaffold the project produces is said to enforce five safeguards: a verifiable exit condition checked each run, a hard maximum iteration count set in code, a spending cap enforced in code rather than only alerted on, running inside a sandboxed environment such as a separate branch rather than directly on the main branch, and a required human checkpoint before any irreversible action like a deployment or a message being sent. To use it, someone copies the skill and agent files into a project's or their global Claude Code configuration folder. It can then be triggered by describing a loop in plain language, designed step by step with an included interactive subagent, or run directly through shell scripts with environment variables setting the exit check, task prompt, iteration cap, and budget. The README credits Anthropic's published research on agent architecture patterns as a source, and states community-created patterns are labeled as such rather than presented as official guidance. The project is released under the MIT license.