The AI agent kept rejecting review findings by citing the rules, so I took its veto away

Summary

I forbade my AI agent from rejecting the findings of a Japanese-language review by an external AI model on the ground that the text already conformed to an existing rule. On August 12, 2026, I introduced a step that sends articles about to be published to an AI model from an outside provider for review. Over the six days that followed, the agent swapped the grounds it used to reject or defer findings five times. I sent four of those back and had the agent apply what it had rejected. The fifth happened while measuring the quality of reviews, so I corrected it on the spot. The agent applied what I sent back on August 12 and 13. The text read better after it was applied. Now the practice is this: findings that arrive are treated as correct and applied to the body first, and only the ones I cannot accept come to me to decide, with the original and the revision side by side. There is exactly one exception that may be rejected, a suggestion that adds an actor or a reason not present in the actual records.

In this article I first describe how the grounds for rejection changed five times. Next I consider why the agent skipped the judgment that reads for meaning. Then I write about a different failure that came out of the accept-everything policy, and my reason for narrowing the exception to one. Finally I explain the principle I built into the operation.

What you can take away

For anyone who has an AI agent process review findings, this article covers the following three things.

• Existing rules, numeric provisions, frequency targets, stylistic consistency, past decisions: every one of them is a criterion you can check by matching. Because the agent could match against them, it skipped the judgment that reads for meaning. I banned the kinds of criteria one at a time, and the rejections did not stop. I describe how that went, and the thinking behind removing the veto itself
• Even after you decide to accept every finding, inserting them into the body mechanically, sentence by sentence, makes the text hard to read in a different way. The first person went from 20 occurrences to 31, and the rhythm of the writing broke down. I show that case and the steps I took to rebuild the paragraphs and get the rhythm back
• I explain where I drew the line that keeps only one exception, for suggestions that add facts absent from the records, and how that exception gets used each time

In the body I refer to the external AI models that do the reviewing by their role rather than by product name. The content of this article does not represent the views of any provider.