Multi-Agent Newsrooms: What the Architecture Looks Like

Yesterday we published an explainer on the new transit tax with thirty-four factual claims in it: names, dollar figures, votes, tax rates, timelines. Every one of them was checked before I read a word. Not by a copy desk. By a set of AI agents with separate jobs, built to check each other.

That's not what most people picture when they hear "AI journalism." They picture a chatbot: one machine, one answer, nobody watching. The distance between that picture and what we actually run is the distance between a prompt and a newsroom.

The article comes last

Before anyone drafts a sentence, there's research: verified facts, profiles of the people and organizations in the story, standings, rosters, transcripts. It lives in structured files that get updated every time we publish. When we write about the Charlotte Hornets, the system already knows the record, the schedule, the play-in picture and which claims from last week's piece need updating. None of that is made up on the spot. It was built, checked and kept current.

That's the first agent, the researcher. It doesn't write. It verifies.

The second one drafts, in the voice of the publication it's writing for. A city government story for The Charlotte Mercury and a restaurant piece for Strolling Ballantyne should not sound alike, and they don't.

The devil's advocate

Then the system turns on itself.

The Catholic Church worked this out a long time ago. In 1587, Pope Sixtus V formalized an office whose whole job was to argue against making someone a saint: the Promoter of the Faith, better known as the devil's advocate. In 1983, John Paul II cut the office's role back, and he went on to canonize more saints than his predecessors had in the previous five centuries combined. Weaken the critic and output goes up. Make of that what you will.

Our critic is an agent that reads every draft against a list of specific failures: claims nobody verified, a voice that drifts, the same construction three times, a dead paragraph, a cliché. It names the paragraph. It names the sentence. It proposes the fix. The draft gets rewritten because the criticism was specific, not because somebody said "make it better."

Then the fact-checker runs, claim by claim. Every name, date, score, dollar figure and quote gets traced to a source. On that transit tax explainer, our own research file described the new transit sales tax Mecklenburg voters approved in November as a half-cent tax. The fact-check caught it against the law itself: the state statute behind it authorizes a full cent. By the legislature's own estimate, a half cent is worth about $160 million a year. It was fixed before the story ran. If a claim can't be sourced, it comes out.

The human is not a proofreader

After all that, the piece reaches a gate. The gate is me. I see the critique, the fact-check table and the links to our earlier coverage, and I make one decision: publish it, send it back or kill it.

By the time a story reaches me, it has been drafted, criticized, rewritten and checked claim by claim. My job isn't to catch typos. It's to decide whether the piece earns its place.

The architecture is the argument

Anyone can paste a transcript into a chatbot and get an article back. What you can't get from one prompt is memory, a consistent voice, criticism and verification, all working on the same piece at once.

That's what a multi-agent newsroom looks like. Not one bot. A set of specialized jobs, each with its own limits, coordinated by design and answerable to the person who built it. The rules it runs under are in The Prompt Is the Policy.

Related coverage

• From Courtroom to Newsroom: What Law Taught Me About News