Loop Engineering Review: Stop Prompting AI Agents — Design the Loop

5,905 stars. 27 days old. One core idea: stop prompting your AI agents — build a loop that does it for you.

That’s the pitch behind cobusgreyling/loop-engineering, an open-source framework that launched on June 9, 2026 and has already crossed 5.9K GitHub stars. It’s not another coding agent. It’s the control system that sits above your agents and tells them what to do — on a schedule, with safety gates, without you touching a prompt.

I spent an afternoon digging through the repo, running the CLI tools, and trying to answer one question: is “loop engineering” a real workflow upgrade, or just prompt engineering with a new label?

The Shift: From Prompting to Designing

Boris Cherny, Head of Claude Code at Anthropic, put it bluntly:

“I don’t prompt Claude anymore. I have loops running that prompt Claude and figuring out what to do. My job is to write loops.”

That’s the mental model. You stop being the person who types prompts and start being the person who designs the system that generates them. It’s the difference between manually watering plants and installing drip irrigation.

Loop engineering formalizes this into five building blocks:

The key insight: none of these are novel individually. Git worktrees have existed for years. MCP has been around since late 2024. What’s new is the composition — wiring them together into a self-running system with guardrails.

Seven Production Patterns (That Actually Ship)

The repo ships seven battle-tested loop patterns, each with its own starter kit and token cost estimate:

• Daily Triage — scans issues/PRs every morning, produces a triage report. Low token cost, safe to start here.
• PR Babysitter — watches open PRs every 5-15 minutes, runs CI checks, comments on failures. High token cost but saves hours of manual review.
• CI Sweeper — auto-diagnoses CI failures, proposes fixes. Very high token cost — only enable at L2 confidence.
• Dependency Sweeper — checks for outdated deps, opens bump PRs. Safe and boring in the best way.
• Changelog Drafter — reads commit history, drafts release notes. Low cost, high value.
• Post-Merge Cleanup — detects stale branches, orphaned comments, leftover TODOs after merges.
• Issue Triage — labels, prioritizes, and routes new issues to the right people.

What I like: each pattern has a confidence level (L1 through L3). L1 is “report only, don’t touch anything.” L2 allows safe automated actions. L3 is full autonomy. You don’t jump straight to L3 — the framework forces incremental trust.

The CLI Toolchain (5 Minutes to First Loop)

The repo isn’t just documentation — it ships five npm packages you can run right now:

# Scaffold skills, state files, and budget config — prints your Loop Ready score
npx @cobusgreyling/loop-init . --pattern daily-triage --tool claude

# Estimate token cost before running
npx @cobusgreyling/loop-cost --pattern daily-triage --level L1

# Audit your loop setup — get a score and improvement suggestions
npx @cobusgreyling/loop-audit . --suggest

# Detect drift between your STATE.md and LOOP.md
npx @cobusgreyling/loop-sync .

# Stateful memory manager with circuit breaker for long runs
npx @cobusgreyling/loop-context --check --ledger run.json

I ran loop-init on a test project. In under 5 minutes, it scaffolded a LOOP.md (the loop definition), STATE.md (persistent memory), SKILL.md (project knowledge), and a budget.json (token cost caps). The Loop Ready score started at 10 (bare scaffold) and climbed to 85 after filling in the project-specific details.

The tooling feels production-minded: circuit breakers prevent runaway token spend, drift detection catches state rot, and everything is inspectable as plain Markdown.

What Makes This Different

The AI tools space is drowning in “yet another coding agent” projects. Loop engineering side-steps that competition entirely. It doesn’t care which agent you use — Claude Code, Codex, OpenCode, Grok CLI, Cursor, whatever. It provides the orchestration layer above them.

This is closer to what Addy Osmani calls “the second wave of AI tooling.” The first wave was about better models and better prompts. The second wave is about system design — how you compose agents into reliable workflows that run without human babysitting.

Two things give me confidence this isn’t just hype:

1. The repo dogfoods itself. validate-patterns and audit workflows run on every push and PR. The repo’s own LOOP.md describes the loops that maintain the repo. You can watch it maintain itself on GitHub Actions.

2. The adoption curve is real. 5.9K stars in 27 days with zero marketing — just a Substack essay and organic word of mouth. The patterns are being forked and adapted for teams using entirely different agent stacks.

Where It Falls Short

Not everything is polished. The MCP server is repo-only (no npm package yet). The documentation is thorough but dense — you’ll spend 20 minutes reading before you understand the primitives matrix. And the seven patterns cover common workflows, but if your use case doesn’t map to one of them, you’re building from scratch.

The biggest gap: loop engineering currently targets solo developers and small teams. There’s no multi-user state management, no team-level audit trail, and no built-in approval workflows for L3 autonomy in regulated environments. If you’re at a 200-person engineering org, you’ll need to layer those yourself.

Bottom Line

Loop engineering earns its stars. It’s not vaporware — the CLI tools work, the patterns are real, and the philosophy (design the system, not the prompt) addresses an actual pain point for anyone running AI agents at scale.

Who should adopt it now:

• Solo developers running Claude Code or Codex daily — the Daily Triage and Changelog Drafter patterns will save you 30+ minutes a day
• Small teams that want to automate rote PR review and CI diagnosis without building custom infrastructure
• Anyone who’s hit the wall with manual prompting and wants a structured off-ramp

Who should wait:

• Teams that need SOC 2 compliance or enterprise audit trails — the framework isn’t there yet
• Developers who only use AI agents occasionally — the setup overhead isn’t worth it for one-offs
• Anyone expecting a GUI — this is CLI-first, Markdown-native, and unapologetically for terminal users

At L1 confidence, loop engineering is low-risk to try. Scaffold a Daily Triage loop on a side project this week and watch it run for 7 days. If it catches even one issue you would have missed, the framework has paid for itself.

Loop engineering is MIT-licensed and available at github.com/cobusgreyling/loop-engineering.