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.
TypeScript agentic-aiai-agentsai-codinganthropicautomationclaudeclaude-codecodexcoding-agentsdevops-automationdevtoolsgithub-actionsgrokllmloop-engineeringmcpprompt-engineering
Dependency Sweeper Week One — Patch Volume and a Verifier Lie
Setup
- Pattern: Dependency Sweeper (L2 patch-only)
- Cadence:
/loop 6h - Denylist:
react,@company/core-auth - Verifier:
loop-verifierin isolated worktree
What Worked
- Patch bumps for transitive lodash CVE closed in 4 hours without human touch
- State file made it obvious what was in-flight vs waiting on human
- Denylist prevented a reckless React minor during a release freeze
What Broke
- Day 3: verifier APPROVED an
expresspatch but did not runnpm ci— onlynpm test, which used cachednode_modules - CI failed on the PR with a peer dependency conflict the verifier missed
- Day 4: loop opened 3 patch PRs in one run — reviewer fatigue
Metrics
| Metric | Week 1 |
|---|---|
| Patch PRs opened | 11 |
| Human merges | 7 |
| Verifier false APPROVE | 1 |
| Token estimate | ~800k |
Lesson
Patch-only is not "risk-free." Verifier must run the same install path as CI (npm ci, not npm test alone). Cap auto-PRs per day (we added max 5 to LOOP.md). One false APPROVE cost more human time than a week of manual patches would have.