A tool-agnostic reference for designing the loop that runs your AI coding agent, instead of you prompting it one turn at a time.
Before you start
Comfortable naming what a skill (a saved how-to Claude draws on automatically) and a sub-agent (a second Claude instance that checks the first one's work) are. New to those words? Start with Add skills and subagents with restraint first. You also need Claude Code running in a real project folder with Node installed, and loop-engineering is a third-party, MIT-licensed tool from one maintainer, not an Anthropic product.
Run the audit tool against your own project. It prints a single Loop Ready score from 0 to 100, a readiness level, and a findings list naming what's missing across automations, worktrees, skills, connectors, sub-agents, and memory (the durable file that survives after the chat window closes). You can't tell what to fix until you see the number.
loop-init creates the starter files for whichever pattern you pick (a daily-triage pattern is the gentlest first loop) plus a budget and run-log file, so you're not hand-writing boilerplate. The command below names Claude; swap in `codex` if that's the agent you're running instead.
Sub-agents and long-running loops can burn tokens fast. loop-cost estimates spend for a pattern and cadence before a single automated run happens. A surprise bill isn't how you want to find out it was expensive. (`--level L1` here is the tool's own L1-L3 readiness scale, unrelated to Agentic Levels.)
The framework's own rollout is report first, then assisted fixes, then unattended. Your first loop should only read your project, write to a memory file, and tell you what it found, nothing else, for at least a week.
Run the same audit command again after you add whatever was missing (a checker sub-agent, a scheduled run, a memory file). Watching the score move from a real 0-100 number is how you know a change actually mattered, not just that it felt like progress.