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
PR Babysitter — Week One (Example Story)
Anonymized production-style narrative for contributors. Replace with your real metrics via PR.
Context
- 4-person team, monorepo, Grok Build TUI + GitHub
- Pain: "can you rebase?" and "CI is red" Slack pings 10–15×/day
Setup
- Pattern: pr-babysitter
- Cadence:
/loop 5mduring work hours - L2 assisted: minimal-fix + loop-verifier, no auto-merge
What Worked
- State file (
pr-babysitter-state.md) stopped duplicate fix attempts — attempt counter mattered - Verifier on stronger model caught 2 over-broad diffs before PR comment
- Loop comments signed
🤖 Loop Engineering — PR Babysitter— reviewers knew what to trust
What Broke
- Day 3: Infinite fix loop on flaky e2e — loop tried code changes 4 times before human killed scheduler
- Fix: Added flake classification to triage; max 3 attempts enforced in prompt
- Day 5: Notification fatigue — bot commented on every run even when nothing changed
- Fix: Only comment when verdict APPROVE or ESCALATE_HUMAN
Metrics (week 1)
| Metric | Before | After |
|---|---|---|
| Slack CI/rebase pings | ~12/day | ~4/day |
| Mean time to first fix proposal | hours | ~25 min |
| Bad merges from loop | 0 | 0 |
Lesson
Start L2 only after 3 days of state-only watching. The loop doesn't know a flake from a regression — you teach it in skills.