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.
The Score Climbed to 100. Then the Budget Burned.
Subtitle: Loop Ready measures scaffold — not permission to let agents write unattended.
You can make an AI coding agent look brilliant for an afternoon.
You can also burn a week of token budget overnight fixing the same flaky test.
This is a field note from dogfooding Loop Engineering — the patterns, starters, and CLIs for designing systems that prompt your agents (instead of you re-prompting every morning).
If you only remember one line:
A Loop Ready score of 100 is a landing checklist, not a green light for auto-fix.
The 15-second dopamine hit
npx @cobusgreyling/loop-init . --pattern daily-triage --tool claude
npx @cobusgreyling/loop-audit . --suggest
On a greenfield or “we have AGENTS.md and chaos” repo, the score can climb from the teens into the 90s once you have:
- a pattern and skills
STATE.md/ loop files- budget + constraints
- a run log
That climb is real. It’s also dangerous if you confuse readiness of the harness with permission to merge.
We shipped a clearer path in v1.6.0: loop-init --with-foundry, loop-audit 1.7 Harness Runtime signals, and loop-gate for mechanical denylists.
Week one: the part that worked
We ran Daily Triage in report-only mode.
No auto-close. No auto-merge. No “just fix main” ambition.
The loop’s job was to produce a human-readable agenda: stale PRs, flaky suites, normalized red builds. Humans acted on the report in standup.
That week was boring. Boring is the goal.
Week two: the part that failed
Day six, we “graduated” to a CI Sweeper on a tight cadence.
We skipped:
- A separate verifier (maker graded its own homework)
- A branch allowlist (feature-branch chaos entered the loop)
- A hard stop on daily tokens (cadence compounds faster than Slack)
Within 48 hours the ledger looked like a runway fire: millions of tokens, symptom PRs, state drift, and one config change that would have broken prod if a human hadn’t caught it.
We killed the sweeper. We went back to report-only. We wrote the failure down so we wouldn’t relearn it as folklore.
(Repo story: score-climbs-then-budget-burns.md · earlier cousin: why we killed CI sweeper.)
The design rule
| Phase | Allowed | Forbidden |
|---|---|---|
| L1 (week one+) | Report, label proposals, draft notes | Unattended code changes |
| L2 (only after L1 is boring) | Fixes in worktrees, gated PRs | Direct pushes, prod-path writes without loop-gate |
| Always | Budget stop conditions, human merge | “The score was high so we turned on everything” |
Loop engineering is not prompt engineering with a stopwatch.
It’s systems design: schedule, skills, state, verification, handoff, budget.
What to run tomorrow morning
npx @cobusgreyling/loop-init . --pattern daily-triage --tool grok
npx @cobusgreyling/loop-audit . --suggest
npx @cobusgreyling/loop-cost --pattern daily-triage --level L1 --cadence 1d
When the score is high and report-only has been boring for two weeks, then consider L2 — and version the loop as a harness (--with-foundry) so you can evolve it without folklore.
Join the loop
- Repo: https://github.com/cobusgreyling/loop-engineering
- Release: https://github.com/cobusgreyling/loop-engineering/releases/tag/v1.6.0
- Show your score: https://github.com/cobusgreyling/loop-engineering/discussions/326
- Ask anything: https://github.com/cobusgreyling/loop-engineering/discussions/327
If you post a score, post a failure too. That’s the culture.
— Cobus