// the find
cobusgreyling/loop-engineering
Practical patterns, starters & CLI tools for loop engineering with AI coding agents. Design systems that prompt and orchestrate agents (inspired by Addy Osmani and Boris Cherny). Includes loop-audit, loop-init, loop-cost.
A pattern library plus a unified CLI (`@cobusgreyling/loop`) for wiring AI coding agents (Claude Code, Codex, Grok, Cursor, etc.) into recurring jobs — daily triage, PR babysitting, dependency bumps, changelog drafts — with a graduated L1 (report-only) to L3 (unattended) rollout model gated by a 'Loop Ready' score. It's for teams already running coding agents who want repeatable, tool-agnostic automation patterns instead of writing the same orchestration prompt from scratch each time.
The CLI consolidates what used to be separate packages (loop-init, loop-audit, loop-cost) into one front door with init/doctor/status/audit/cost subcommands, which is a real ergonomics win over juggling N npm installs. The Loop Ready score explicitly weights recent run history over static files on disk, directly targeting the common failure mode where a 30-day-stale STATE.md gets mistaken for a working unattended loop. Patterns are defined tool-agnostically with separate worked examples per agent (Claude Code, Codex, Grok, Cursor, Windsurf, Hermes) rather than hardcoding one vendor's prompt format. docs/failure-modes.md, anti-patterns.md, and a stories/ directory that collects failures alongside wins are first-class, not bolted on after launch.
The repo checks in its own growth-hacking collateral under docs/distribution/ — an arxiv paper thread, a Show HN draft, fork outreach scripts, social media drafts — which reads less like project documentation and more like a marketing playbook shipped alongside the code. scripts/issue-bodies/ holds dozens of pre-written issue templates split by tool and pattern combination (cursor-ci-sweeper-example.md, windsurf-dependency-sweeper-example.md, hermes-pr-babysitter-example.md...), which looks like manufactured 'good first issues' rather than things that came up from real usage. The ecosystem has sprawled into 10+ separately released sub-packages (loop-audit, loop-context, loop-cost, loop-gate, loop-mcp-server, loop-sandbox, loop-swarm, loop-sync, loop-worktree) plus five more companion repos the README itself says not to touch yet — that's a lot of version-alignment surface for a tool whose pitch is reducing orchestration complexity. The README's authority claim ('this repo is the community reference' reviewed in an arxiv paper) is doing more work than the patterns themselves to establish credibility, which is worth discounting until you've tried a pattern against your own repo.