Skip to writing
Peter Cellino — The Working Journal

Apr 7, 2026

What Agentic Control Actually Means

Most people hear "AI-generated news" and picture a machine with no editor. James Watt tackled this problem in 1788 with two spinning iron balls. Control isn't supervision. It's a governor.

Most people hear "AI-generated news" and picture a machine with no editor. Wrong picture. The question was never whether a machine has an editor. It's whether somebody designed the machine to check itself.

The bottleneck was never the writing

The traditional version of editorial control is a person. An editor reads the draft, catches the bad sentence, flags the claim nobody checked and sends it back. That works when you have three reporters and a city editor with a red pen and enough coffee to get through the budget recap before deadline.

It doesn't work when you're publishing across six sites with a staff of one human and a set of AI agents that can draft faster than you can read.

Here's what most people get wrong about scale: the bottleneck was never the writing. It was the control. An AI agent can turn out a 700-word council recap in about ninety seconds. If the only quality check is "Peter reads it before it goes live," you've automated the production and left quality control in the era of the red pen. You've traded one bottleneck for another.

The governor

James Watt had this problem in 1788. A steam engine left to itself runs too fast or too slow, and nobody can stand next to it working the valve all day. So he fitted his engines with a centrifugal governor: two iron balls on a spinning shaft. When the engine sped up, the balls swung out and eased the throttle closed. When it slowed, they dropped and opened it. The engine kept itself in line.

Eighty years later James Clerk Maxwell wrote a paper about it called "On Governors," and control theory, the study of machines that correct themselves, is usually dated from there. When Norbert Wiener named the broader field in 1948, he called it cybernetics, from the Greek word for a steersman, the same root that gives us "governor."

That's what control means here. Not a person standing next to the engine. A mechanism built into it.

I've written about the agents that research, draft, criticize and verify each other's work. The principle is what matters: by the time an article reaches me, it has been checked against seventeen editorial criteria, rewritten to fix what failed, and verified claim by claim against a sourced table. My job at the gate isn't copyediting. It's a judgment call: publish it, send it back or kill it.

You get what you design for

You get the quality you design for, whether the writer is a person or an agent.

The technology supports structured criticism, revision, source checks and voice. What it doesn't support is laziness. Deploy an agent with no pipeline, no gate and no fact-check, and you get exactly the quality you invested in, which is none.

The question was never whether AI can write a good article. It's whether the person deploying it cared enough to build the system that makes a good article the default instead of an accident.

Control isn't supervision

Supervision is a person watching a machine. Control is a machine built to watch itself.

Every piece we publish goes through the same pipeline, the same gate and the same verification table. The system doesn't get tired. It doesn't skip the fact-check because the meeting ran late. It's built so that a detail nobody can source gets cut, even when the paragraph sounds good.

I get tired. So I built a governor. The rules it runs on are in The Prompt Is the Policy.


Related coverage