A tool waits for your hand. An agent doesn't. You hand it a goal instead of a grip, and it works out the rest, the steps, the order, the moment to stop. The whole craft is in that handoff: what you let move without you, and how close you stay to the wheel.
For two books the machine talked. You asked, it answered, and you decided what to do with what it said. Then it learned to act, to take a step, look at the result, and take another, running a small loop of its own while you weren't watching. The thing on your screen at the end wasn't a reply. It was a result. That shift, from a system that answers to a system that acts, changes what can go wrong, how fast, and whose name is on it.
A More Beautiful Agent is a working guide to building agents you'd actually trust unwatched. It treats agent design as the oldest human skill there is: handing off real work without losing the thread, the thing every manager, parent, and delegator has always had to learn, now aimed at a machine that takes you at your word and runs faster than you can supervise.
At its center is a simple framework, LOOP, built on the four movements every agent runs:
Lay the groundwork: the goal, the tools, the limits, before anything spins.
Operate: how the agent plans and acts.
Observe: how it reads results, catches its own errors, and knows when it's done.
Preserve oversight: the checkpoints, the trail, the off switch that keep the run yours.
The two inner movements belong to the machine. The two outer ones stay with you. Get that boundary right and an agent extends your reach like nothing before it. Get it wrong and you've handed the keys to a driver who doesn't know your address.
From there the book gets practical: a hands-on gym, worked sessions built and run start to finish, a tour through fifteen professional fields from software and medicine to finance, law, and the physical world, a field guide to the dozen ways agents fail (the runaway loop, the confident liar, the over-permissioned agent that deletes the database), a clear-eyed chapter on agent security and prompt injection, and an honest look at what's coming when agents start calling other agents.
It's written in plain language for the person who has to ship, not the person who wants to admire the theory. No hype, no doom. Just the craft of deciding what a capable machine should be allowed to do on its own, and how you'd know if it overstepped.
This is the third and final book of a trilogy that runs from prompting (A More Beautiful Prompt), through context (A More Beautiful Context), to agents. You don't need the first two to start here. The machine has stopped asking permission. This is the book about what to permit.
