MERIDIAN runs a pipeline of six specialist agents that plan and produce marketing content. Five of them make things. The sixth makes nothing at all, its only job is to read what the others produced and object. We call it the critic, and it has quietly become the part of the system clients ask about most.
The critic gets the brand guidelines, the campaign brief, and the draft. It returns a verdict with reasons: this claim is not supported by the brief, this phrasing has appeared in the last four posts, this tone does not match the audience. If it objects, the draft goes back, even if a human already marked it as fine.
Why it catches what people miss
Humans reviewing a queue of drafts get tired and get generous. By the fifteenth piece, “good enough” drifts downward and nobody notices, because the drift is slow. The critic reads the fifteenth draft exactly the way it read the first one. That consistency, not intelligence, is what makes it useful.
It also changes the politics of review. When a colleague rejects your work, it costs something socially. When the critic does, nobody takes it personally, the team argues with the objection on its merits, which is precisely the argument worth having.
“The critic is not smarter than the people it overrules. It is just never tired, never polite, and never in a hurry.
The audit trail matters more than the verdicts
Every objection is logged with its reasoning, and every override of the critic is logged with a human name attached. Six months in, that log is the best record we have of what the brand’s standards actually are, not the guidelines document, but the accumulated history of what got rejected and why. Several clients now read the objection log the way they used to read performance reports.
If you are building a multi-agent pipeline and can only afford one more agent, make it the one that says no.