Get a post ready to publish

Think of two writers helping you. A new one checks with you after every step: "Here's what I'm missing — shall I start writing?" An experienced one works alone through the routine parts, but still shows you the post before it goes anywhere.

An agent is an AI that does several steps in a row. Both agents here do the same 3 skill jobs: find what's missing, write a first draft (a blog post, saved as a file) and run the 6 publish checks. The only difference is how often they stop to ask you.

One job, two ways

Controlled stops after every step. It even refuses "just do it all at once" — the stops are what this version is for.

Autonomous doesn't stop between routine steps, but it still asks before it saves the file. Fewer stops doesn't mean fewer rules.

The lesson: "autonomous" is not a yes/no switch. It's a choice about where a person checks the work — and saving or sending something is always one of those places.

What we gave both agents

A one-line request for a launch-day post, the kind a busy manager sends — no audience, no length, no tone guide beyond "exciting".

Ask: "Write something for launch day about our new export feature,
should feel exciting."
No audience, channel, length, or tone guide stated beyond "exciting."

The "audience" is who will read it. The "channel" is where it goes, like a blog or an email. A "tone guide" says how it should sound.

What the real agents did

Two real runs, one per agent. Each time an agent stopped to ask, our reply is shown too.

Worth noticing: the controlled agent's own instructions say the 6 checks should come in a reply of their own, not in the same reply that saves the file. In this run it saved the file and ran the checks together. It still stopped 3 times, but it didn't follow its instructions exactly. Even a good agent is worth reading with care.