A writer agent and a critic agent: an n8n loop that ends on a rubric, not on fatigue

A writer agent and a critic agent: an n8n loop that ends on a rubric, not on fatigue#

The idea is simple: one model writes, a second reads critically, the first revises. The trouble is that the naive version ends one of two ways. Either the critic starts nodding along after round two because "the draft improved", or the loop spins forever on cosmetics.

This n8n workflow is built around its stop conditions. The prompts come second.

Loop state in one object#

Everything the loop carries between rounds sits in one object: round number, current draft, latest critique and score history. The node at the head of each round changes nothing; it just passes that state on. That keeps the edge back to the start of the loop easy to follow in the editor.

The writer may disagree#

The writer must address every critique point. If it deliberately disagrees with one, it says so in deliberate_deviations instead of quietly ignoring it. Missing data gets a placeholder such as [FIGURE NEEDED: Q3 revenue] and goes into needs_from_human rather than an invented number.

The critic scores, it does not rewrite#

The critic gets the brief, the house rules and the rubric. It splits its points into must_fix (breaks the brief or the rules) and would_improve (preference). The prompt says outright:

Do NOT inflate the score across rounds to let the loop finish.

Approval needs two things at once#

const meets_bar = !isNaN(score) && score >= threshold && mustFix.length === 0;

A high score with an open must_fix point is not an approval. The rubric and the blocking list are separate signals, and the stricter one wins.

Three ways out of the loop#

The round limit guarantees termination whatever the two models agree between them. The sheet log records not only the result but the exit reason, the score history and what the text still needs from a person.

Configuration#

Limitations#

The writer and the critic are the same kind of model. They share blind spots, so the critic may miss a mistake the writer made for the same reason.

Each round is at least two model calls, and because the full draft is passed along, cost grows with text length.

Output quality depends on the rubric. A vague rubric gives vague critique.

The workflow does not check facts. It only makes sure the text introduces no numbers that were not in the brief.

When it does not make sense#

For one short text, editing it yourself is faster. The loop makes sense for repeatable texts with clear rules: internal notes, product descriptions, replies that follow a pattern. We write about drawing boundaries between agents in every agent needs to be told what it does not do. When texts like these should be produced inside your own systems, that is work for AI process automation.