We Published Our Own Patterns Without Our Own Check

On the morning of October 28, 2026, I told my agent to publish our new catalog of AI Patterns: the named, repeatable ways our organization of people and agents works. "Publish all so I can see them on the site," I wrote. Minutes later, 48 pages were live on this site: 47 patterns and the page that lists them.

None of them had been through our Compliance Agent. Every article, every AI Definition, and every post on X passes that review before readers see it. The catalog was a new kind of page, and the check lived inside the old kinds. So the new kind went around it.

What happened

Two things caught it, and both depended on me paying attention. First, I noticed that one pattern named the people who review our knowledge. A pattern is meant for any organization, so I asked that every pattern use roles, not our names. Then I asked a question: had the Compliance Agent reviewed all this?

My agent answered plainly: no. It ran the review on every page as readers saw it. Three of the 48 pages passed. The other 45 drew 109 findings. Many were real. Some pages pointed to internal documents that readers cannot see. Some called a pattern our own when it is established practice elsewhere, such as keeping one source of truth for each fact. Some marked as Running practices that had started that week. Many citations fell short of our own standard.

Some findings were not real. The Compliance Agent had been given only one of our three books, The Knowledge Core. When a page cited Building Agentic Organizations or Building AI Agents, it read the citation as the wrong title for that one book, and the chapter as one that does not exist. The reviewer had a blind spot of its own.

I said yes to taking the catalog down. The pages were public for about eleven minutes, and the correction is on our Corrections page.

The irony

The catalog that failed contains the patterns that should have stopped it. One is Systems Enforce, Instructions Guide. My book puts it in one line: "If a boundary matters, put it in a system."¹ Another is Exact Text, Then Yes. Knowing the patterns did not stop us from skipping one. Writing a rule down is not the same as building it into the work.

What it teaches

Oversight must not depend on attention. The Knowledge Core says so in as many words: "Oversight does not depend on anyone's attention."² Here it did. A person in the middle worked, but only because I happened to ask. On a busier day, I would not have.

"Publish" means publish what passes. My yes was a decision about the catalog, not a substitute for the review. Nothing told my agent the review applied to a new kind of page, so it did not run one. The fix is not a more careful agent. It is a step that cannot be skipped.

Reviewers need the whole record. A reviewer given only part of the sources raises false alarms, and false alarms teach people to stop listening. A review is only as good as what the reviewer is given.

Agents should point out holes. When I asked, my agent said no, ran the check, and brought me the results with a recommendation. That is what we ask of every agent: say what you could not do, rather than hide it.³

Correct plainly. The pages are down, the mistake is on our Corrections page, and this piece is the account.

The fix: check at the door

We moved the check. A pattern is now published through one step that refuses any page whose exact text, as readers would see it, has not passed the Compliance Agent's review in that version. The reviewer is given every book the page cites. A new kind of public page goes live only with the same check at its publishing step. We call it Check at the Door: the review sits where publishing happens, not inside each kind of work. It is Early: we began it the day this happened.

The strongest objection

A reasonable leader would say this is overbuilt. The pages were up for minutes, a person caught the problem, and no email or post went out with them. Why add machinery when a careful person and a careful agent are enough?

Because careful people are not always paying attention, and the cost of the gate is a few minutes of review. My book's rule is to keep two things apart: "what an agent may do, and how closely it is watched."⁴ My agent may be trusted with more each month. How closely its work is checked should not change with it.

A question for you

Where in your organization does a check live inside one kind of work, so that the first new kind of work will go around it?