Sunday, March 29, 2026
Charlotte, NC|
Notes

Multi-Agent Newsrooms: What the Architecture Actually Looks Like

Peter Cellino explains the multi-agent architecture behind Mercury Local — how specialized AI agents research, draft, critique, and fact-check each other's work before any human reads a word.

Peter Cellino· Publisher
||3 min read

We published a transit explainer last week with sixty-five factual claims — names, dollar figures, board votes, tax rates, project timelines. Every one of them verified. Not by a copy desk. By a system of specialized agents that research, draft, critique, and fact-check each other's work before any human reads a word. That is not how most people picture AI in journalism. Most people picture a chatbot. One machine, one output, minimal oversight. The distance between that image and what we actually built is the distance between a prompt and a newsroom.

The Article Is the Last Thing the System Produces

Before a single sentence gets drafted, there is research — verified facts, entity profiles, standings, rosters, source transcripts. That research lives in structured files that get updated every time we publish. When a piece runs under Jack Beckett's byline about the Charlotte Hornets, the system already knows the record, the schedule, the play-in standings, the three-point tracker, and which claims from last week's piece need updating. None of that is generated on the fly. It was built, checked, and maintained as a running body of institutional knowledge.

That is the first agent: the researcher. It does not write. It verifies.

Specialized Roles, Not General Intelligence

The second agent drafts — but not generically. It drafts inside a specific editorial identity. Beckett's voice is encoded in a document the drafting agent reads before writing a word: vocabulary, sentence rhythms, words he uses and words he never uses, how he opens a piece covering a zoning meeting versus a playoff game. John Speedway's voice is a different document entirely — a different register, different reference points, a different relationship with the reader. This is not a style toggle. It is a complete editorial personality. The difference between a generic AI article and one written inside a voice profile is the difference between a form letter and a column.

Then the system turns on itself. A critique agent reads the draft against a checklist of specific failure modes — voice drift, unverified claims, repeated constructions, dead paragraphs, cliché. It names the paragraph. It names the sentence. It proposes fixes. The drafting agent rewrites, and the piece improves not because someone said "make it better" but because the critique was structural and specific.

Then the fact-checker runs. Claim by claim. Every name, date, score, dollar figure, and quote gets traced to a source. In that transit explainer, an early draft described Charlotte's transit tax as a "half-cent sales tax." The fact-check caught it — the PAVE Act authorized a full one-cent increase, not half. That distinction is worth roughly $165 million a year in public revenue. It got corrected before I ever saw the piece. If it can't be sourced, it gets cut. The system treats unverifiable claims as defects, not approximations.

The Human in the Loop Is Not a Proofreader

After all of that — research, draft, critique, verification — the article reaches a gate. The gate is me. I see the critique notes, the fact-check table, the cross-links to prior coverage. And I make a decision: publish, revise, or discard.

This is the part most people miss. The human isn't reviewing output from a single machine. The human is reviewing the product of a system where multiple specialized functions have already interrogated each other's work. By the time the article reaches my desk, it has been drafted in voice, critiqued for structure, rewritten with fixes applied, and verified claim by claim. My job isn't to catch errors. My job is to decide whether the piece earns its place in the publication.

The Architecture Is the Argument

Anyone can paste a transcript into a chatbot and get an article back. What you cannot get from a single prompt is institutional memory, editorial identity, structural self-correction, and verified accuracy — all operating simultaneously on the same piece.

That is what a multi-agent newsroom looks like. Not one bot. A system of specialized roles with distinct constraints, coordinated by design and governed by the person who built it.

That's not a shortcut. That's a newsroom that checks its own work.

Peter Cellino

Publisher

Publisher of The Charlotte Mercury and its family of hyperlocal news publications.

More in Notes