Just for you (polished, and reduced for 2026 reading format):

Chapter Four: An Agent Decides; a Workflow Does Not

Here is the sharpest distinction in the book, and most of your daily frustration dissolves once you can see it.

An agent is the thing that decides. A workflow is the thing that does not.

Almost everything that goes wrong between you and the machine is a confusion about which of the two you are looking at.

There is a clean test for it. If a step runs every time, it is not a decision, it is a workflow, however clever it looks. The deciding has gone out of it, and what remains would be more reliable as plain code.

With that test in hand, look again at the tasks that keep going wrong. You will usually find one of two mistakes. Either you handed a real decision to a workflow, so it cannot adapt when the situation shifts, or you asked a mind to grind through work a workflow should have owned, so it is slow, expensive, and prone to drift. The frustration was never the machine being bad at its job. It was the job being given to the wrong kind of worker.

The repair is to decide what truly needs deciding, and to make everything else deterministic. A good workflow then behaves like a switchboard. It runs the fixed steps on its own, and it calls a mind only at the step where a real decision lives, handing that mind enough context to decide well and taking the answer back.

It also routes the work to the mind that fits. The deterministic parts go to tooling, which is fast and exact: the editor performs the refactor, the script gathers the data, the check runs. The decision goes to whichever mind belongs there, sometimes the agent, sometimes you. The agent can even reach back for you on its own, opening an issue and naming you when it hits the one judgment only a human should make, then carrying on once you have made it.

This is why you so rarely need to drag a mind across a boundary it is weak at. Each intelligence works from its own side, and the workflow carries the work and the calls between them. You stop trying to make the machine into a reliable tool or a trustworthy judge, and you start placing each piece of the work where it was always going to succeed.

How much of the work that keeps frustrating you never needed a decision at all?

Please go wild with the critics