Loop Engineering: How to Stop Prompting Agents and Start Designing the Systems That Do It for You (Paperback) (ISBN-13: 9798184582740)

You have been getting good at working with AI agents. You know how to write a precise instruction, how to load the right context, how to read what comes back and ask the right follow-up. Your outputs are consistently useful. Your process is disciplined. By any reasonable measure, you are ahead of the curve.

And you have started to notice the ceiling.

The work that fills your week does not run on a single turn. The recurring research, the code reviews, the document processing, the synthesis tasks that happen again on Monday the same way they happened last Monday. You are holding the tool through every step of every run. The agent produces nothing without you there to initiate the next step. Your capacity is bounded by how many sessions you can open and how many hours you have to run them.

Loop engineering is what sits above that ceiling.

A loop is a system that finds the work, hands it to the right agent, evaluates what comes back, records what happened, and decides what happens next. It runs on a schedule. It applies a verification gate before logging anything as complete. It persists state between runs so each cycle builds on what the previous one produced. It has hard cost controls that stop it when something goes wrong. And it runs while you are doing something else entirely.

This book teaches you to build one.

What you will build:

By the end of this book, you will have one complete loop running real recurring work from your own practice. Not a demonstration. Not a toy example. A loop designed for a specific task, built to a complete design document that precedes every line of configuration, running with verified outputs and a budget it cannot exceed.

How the book is structured:

The book follows the ACT framework: Architect, Construct, Trust.

Architect is Block 1 - the design work that happens before any tool is opened. Goal statements with success criteria a verification gate can confirm. Memory schemas written before the first trigger fires. Subagent scope definitions narrow enough to produce consistent outputs. Design decisions made on paper so they are not discovered at runtime cost.

Construct is Block 2 - building the loop in dependency order. Automations configured correctly for the work type and the failure modes. Subagents designed with clean handoff specifications. Verification gates that catch what the agent's self-assessment misses. Hard cost brakes, iteration ceilings, no-progress detectors, spend caps at three time granularities, in place before the first trigger fires.

Trust is Block 4 - the ongoing practice that determines whether the loop stays a tool the practitioner controls or becomes a system the practitioner merely observes. How comprehension debt accumulates silently. What the practitioner must be able to answer about their running loop to remain the engineer. How to detect when a loop has drifted from the work it was designed to do, and how to update it without invalidating what it has already produced. When to stop a loop deliberately and how to do it gracefully.

Between Construct and Trust, Block 3 presents three complete reference loop designs, code review, research synthesis, and task processing, each with a full anatomy, goal statement, memory schema, subagent structure, verification gate, and cost model at both Sonnet and Opus pricing.

Who this book is for:

Practitioners who use AI agents for serious recurring professional work and have begun to notice the ceiling of the prompt-by-prompt model. Technical enough to understand configuration. Not required to be a software developer. If you have spent meaningful time working with AI agents and you find yourself wishing they could just keep going after you stop watching, this book is about the design that makes that possible