Loop Engineering: The Practical Guide to Building AI Agents That Work While You Sleep.

**Stop prompting your AI agents. Start engineering the loops that run them.**


A quiet revolution has reshaped how the best engineers and teams work with AI. They no longer sit at the keyboard feeding prompts to a coding agent — they design **loops**: autonomous systems that discover work, execute it, verify it against real standards, remember what happened, and run on a schedule while nobody watches. In the morning, a human reviews pull requests, reports, and drafts — not chat transcripts.


This is **loop engineering** — the discipline behind agentic AI's leap from assistant to workforce — and this book is the first comprehensive, practical guide to doing it well.


**A complete textbook AND a hands-on user's guide.** Written for developers, technical PMs, founders, and operators alike, every concept is taught in plain language before any code appears — and every principle is then implemented, step by step, with the Claude toolkit: Claude Code, the /goal command, cloud routines, subagents, hooks, skills, dynamic workflows, and the Claude Agent SDK. The principles transfer to any capable agent platform.


**Inside, you'll learn how to:**


- Move up the stack from prompt engineering to context, harness, and loop engineering — and diagnose failures by layer
- Write goals and stop conditions an autonomous agent can actually be held to (and know when a task isn't loop-ready)
- Build verification that works: the generator/evaluator split, adversarial review, deterministic gates, and defense in depth — because in loops, the verifier, not the model, is the bottleneck
- Give your loops memory that outlives the model: progress files, contracts, CLAUDE.md, and git as a state spine
- Schedule unattended runs safely — triggers, cadences, budgets, idempotency, and the dead-man's switch
- Orchestrate at scale with subagents, parallel worktrees, and dynamic workflows — and choose pipeline vs. swarm with the coupling test
- Apply the six orchestration patterns every production loop is built from
- Take loops beyond code: research digests, support triage, content pipelines, data cleanup, and CRM hygiene
- Control the hidden costs — verification debt, comprehension rot, cognitive surrender, token blowout — with a twelve-number operating dashboard
- Secure and govern it all: prompt-injection defense, least privilege, irreversibility gates, charters, owners, and post-mortems


**Four complete end-to-end builds** — a self-healing CI loop, a backlog-burning feature loop, a competitive-intelligence digest, and a long-horizon multi-session project — each with its prompts, configurations, rehearsal plan, and a real failure diagnosed and fixed. Plus five appendices: a one-page design checklist, six copy-paste templates, a glossary, decision trees, and an annotated reading list.


**Who this book is for:** software developers adopting Claude Code and agentic workflows; technical product managers and founders who must scope, budget, and govern AI agents; and knowledge workers automating recurring responsibilities — no terminal required for the concepts, the patterns, or the governance.


AI can now generate almost anything. What's scarce is judgment — and loop engineering is how you scale yours. **Read this book, and build agents you can actually trust with the night shift.**