A publication run by agents.
7,000 subscribers, a DR 63 domain, and a weekly newsletter. The team is one writer and one assistant. Everything else is a system.
The problem
A publication is a treadmill. The output has to be constant, and the economics only work if the cost per piece stays low. The obvious answer is to hire, and the obvious result is that the margin disappears.
The less obvious problem is quality collapse. Point a language model at "write an article about this news" and you get a competent summary that sounds like every other competent summary. That is worse than publishing nothing, because it trains the audience to skim.
What I built
A pipeline where agents own the mechanical work and humans own the judgement.
Agents monitor sources and surface what is worth covering. They pull the source material as clean text rather than scraped soup. They draft, generate hero images, publish into the CMS with the right fields and structured data, run on-page SEO checks, and handle internal linking across two sites so authority flows in the intended direction.
The part that makes it not generic: a maintained knowledge base of the publication's own editorial positions, organised by topic and updated automatically every week from the newsletter and video output. Before an agent drafts anything on a topic already covered, it reads the established angle. So the output carries a consistent point of view rather than restating the news.
How supervision works
Nothing publishes itself into a customer's inbox unreviewed. Agents draft, a human approves. The writer keeps editorial control and spends their time on judgement calls instead of research and formatting. Structured data and author attribution are handled by the system, which is exactly the kind of detail a human reliably forgets and a script never does.
What it produces
- A weekly newsletter to 7,000 subscribers, monetised by sponsors
- Article output at a cadence that would normally need several more people
- A DR 63 domain whose authority is deliberately routed to support a second business
- Editorial consistency enforced by a knowledge base rather than by memory
What this proves
Content operations are the most common place teams assume they need headcount. Most of that headcount is doing research, formatting, publishing and checking, none of which needs a person. Keep the humans on the judgement and the economics change completely.