Stop prompting your agent — build the loop that does it for you

What it does

This is a book, not code — a plain-language guide to “loop engineering,” the term that surfaced in June 2026 when Peter Steinberger, Boris Cherny, and Addy Osmani all named the same shift within a week. The core idea: stop being the person who prompts the agent, and instead design the system that does it for you. It’s 9 sections across 4 parts, available as free PDFs in both Chinese and English.

The interesting bit

The framing is architectural: loop engineering sits one floor above harness engineering. The harness equips a single agent run (tools, done-criteria); the loop is the outer system that runs on a timer, spawns helpers, verifies work, remembers what it did, and decides what’s next. Part 3 is where it earns its keep — real loops like Addy’s morning triage and Stripe’s Minions, plus four honest costs: verification debt, comprehension rot, token blowout, and cognitive surrender.

Key highlights

• Free PDFs in Chinese (v261020, October revision) and English (v260615), MIT-licensed
• Covers the prompt → context → harness → loop stack, and why an AI can’t grade its own code
• Grounded in first-hand public sources: Osmani’s founding post, Anthropic’s harness-design blog, Stripe’s Minions, official Claude Code / Codex docs
• October Chinese edition adds /goal, /loop, /usage tested on Claude Code 2.1.287 and Codex CLI 0.163.0, with errata from June (including a misattributed quote)
• Also readable chapter-by-chapter on WorkBuddy’s knowledge base

Caveats

• The English edition is still the June first edition — it lags the revised Chinese version
• Written the week the term emerged, so parts read as rapid-response journalism rather than settled practice; the README itself says to defer to official docs as tools evolve

Verdict

Worth 20 minutes if you’re driving Claude Code or Cursor prompt-by-prompt and suspect there’s a level above that. Skip it if you want code — this repo is a PDF distribution, not a tool.