A More
Beautiful
Agent
The Art of Human-Centered Agent Design
A tool waits for your hand. An agent doesn't. You hand it a goal instead of a grip — and the whole craft lives in that handoff: what you let move without you, and how close you stay to the wheel.
What it's about
Handing over the keys.
The first two books were about what you say and what the model sees. This one is about what it does. An agent doesn't wait for your hand on every step — you hand it a goal, and it works out the sequence, the tools, and the moment to stop.
That handoff is the whole subject: how much to let move without you, how to keep a hand near the wheel, and who stays accountable when a machine acts on your behalf. It is the difference between motion and progress — wider than it looks — and the human judgment that keeps autonomy honest.
The framework
The LOOP framework
Designing reliable AI agents — four movements that turn a goal into trustworthy action, split by the reversibility line.
Before anything runs
Define "done" in observable terms and grant the minimum set of tools needed to finish the job.
Plan, then act
The agent sequences its plan before acting, so motion leads to actual progress.
Verify outcomes
Check real-world results and catch quiet errors rather than trusting a successful "tool call."
The hand on the wheel
Checkpoints at irreversible steps and an off-switch keep the human the ultimate author of conduct.
Where to draw the line: let the agent act on its own for cheap, reversible steps — but require your sign-off before anything permanent
“Trusting someone to act is its own kind of work — and letting go well is harder than holding on.” From the dedication