What the agent cannot see

I build our curriculum with an AI agent. It writes a month of lessons in an afternoon, and it will cite a book we do not own — not because it is careless, but because it has never been in the room. Here is the gate I built between the plan and the child.

A month of lessons for my daughter is not something I write by hand any more. I work with an AI agent. I decide what the month is for, it drafts the days, and between us we get four weeks of real material in an afternoon rather than a fortnight.

The thing to understand about that arrangement is what the agent has access to. It can read every file I have — the curriculum, the assessments, the record of what she has already done. It cannot walk into the study and look at a shelf.

So it reads the book catalog. And the book catalog is a list of books we have bought.

Four books that were not there

Month two came back drafted and good. The Korean reading anchor was a detective book called 나도 탐정이다 — the whole Korean block for the month pointed at it, chapter by chapter.

We do not own that book.

Three others were in the same state. A brain-development workbook, a Japanese number-puzzle book, a maths series. All four written into a month of teaching, none of them in the house.

Nothing had gone wrong with the agent. It had done exactly what I would have done with the information available to it: found the file listing our books, and believed it.

Why a catalog lies

A purchase record and a shelf are different lists, and I had never noticed because I have eyes.

A catalog says we decided to have this. It does not know about the book that was ordered and never arrived, the one bought and lent to a cousin, the one bought for the nine-year-old and long since absorbed into the twins’ pile, the one we meant to buy and recorded our intention to. Every entry is true about the past and silent about the present.

Standing in the room, none of that matters. Working from a file, it is everything.

This is the general shape of the problem when you build with an agent: it is not going to invent a book. It is going to trust a document that answers a slightly different question than the one you needed answered. And it will do that confidently, in fluent prose, across four weeks of work.

I had already been bitten by the same class of thing elsewhere in the same build. A status file said the project folder was not under version control; it was. A profile said Aini was eight years and seven months old; she was eight years and eleven months, because the file had been written four months earlier and nobody ages a document. Small, plausible, wrong.

The fix is not care

The obvious response is to promise closer attention next time. That does not work, because inattention was not the cause. You cannot fix a bad source by reading its output more carefully.

So the rule now sits before the writing rather than after it.

The month gets drafted at a high level first, and that draft has to produce a materials list — every book, every workbook, every instrument, every physical thing the month will touch. That list comes to me before a single day file is written.

Then I take photographs of the actual shelves and send them back. Each title is marked confirmed or uncertain against what is visible in the pictures, and only confirmed books may appear in a lesson.

And then the part that costs me the most time and is worth all of it: for every confirmed book, I open it and read the table of contents out. Chapter five, electricity, page 47. Chapter six, magnetism, 57. Not inferred from the title, not looked up online. Opened, and read off the paper.

The gate is absolute, because a soft version of it is the same as not having it:

Can you start week files before book confirmation? No. Not even a draft. The teaching script cites specific pages; a placeholder creates invisible debt that gets taught wrong.

A draft with “page TBD” in it is worse than no draft. A TBD is a promise to check that nobody keeps, and it survives into the version I am teaching from at eight in the morning with a child sitting next to me.

What the audit actually produced

Photographs went back and forth for an hour. The results were unglamorous and exact.

The Goosebumps volumes were there — eight of them — so those became the Korean reading anchor instead. The physics book was there, and its contents came back as real page numbers: electricity at 47, magnetism at 57, the iron filings experiment at 64, light at 69. The music book was there: posture on pages 7 and 8, the first exercise on 10, Twinkle Twinkle on 11.

Then the removals. The brain-development workbook came out of the plan and the same reasoning got done out loud instead. The number puzzles came out with no replacement, because that thinking was already happening in conversation. And 나도 탐정이다, which had been the spine of the whole Korean month, was replaced by the books actually in the house.

The month lost nothing. What it gained was the property that every page number in it was real.

The part worth taking

If you are building anything with an agent — a curriculum, a meal plan, a trip — the failure will not look like an obvious error. It will look like a good plan resting on a document that was true once.

So the rule generalises past books: a plan may only cite things a human has physically confirmed, recently. The experiment that needs a material you think is in the garage. The recipe that assumes a pan. The outing that assumes a place is still open.

The agent is not the weak link. The weak link is every file that describes the world as it was, and the fact that nothing in the file will tell you which parts have gone stale.