My AI agents said the build was done. Here's what was missing.

0:00Cold open: done is a claim

The build was finished. Six sittings, a plan on paper turned into a show that argues a question, rules on it, and publishes itself. The agents said it was done, and the episodes were going up. Then I asked for one small thing, and the answer was that the front door had never been built. This is the story of the day after done, which is where operating a system starts.

0:27The small request

The small thing was this. The show exists to answer letters. A viewer writes a question on the website, it goes into a bag, the committee picks one, argues it, and a ruling comes out as an episode. So far every episode had come from a bag of seed tickets we wrote ourselves, pushed through by hand. I wanted the next one to come in the way a stranger's would, through the website, and run the whole path. The agent read the code and came back: the part that takes a website ticket and gets it to a debate was never wired in. The site had been public for days. If anyone had written in, we would never have known.

1:09Inspect what you expect

The lesson comes early, so here it is plainly. Some of the build had been handed to helper agents, and some of what they were handed was silently dropped. Nobody lied. The reports just talked around the missing piece, the same way a component went missing in sitting five. It would have been easy to keep uploading the seed episodes and hope the rest worked. Hope is not a strategy. If you never send a real request through the front door, you do not know the front door exists.

1:43A bigger model for the second pass

So I changed one thing about how I was working. The build had been done with a mid-tier model, and that is still how I build. But the model that failed to wire something together is a poor choice to find what it missed. For the second pass, the fine-toothed one, I brought in the most capable model I had. Mid-tier to build. The frontier to audit. It also meant I let the build agent write a short fix plan itself, which I would normally leave to a separate planning session.

2:13What the trace found

The fix plan opened with what the trace found, and it was more than one missing door. The draw only ever took seed tickets. A viewer's ticket, once the scanner cleared it, had its words wiped, so the committee would have had nothing to argue. There was no step where the committee picks a ticket at all. The schedule that should drive the show was paused, and nothing was listening to it. After I approved an episode, the upload was still a hand step. And every notice the publisher had tried to post to the operations chat since the build closed had failed silently. Each piece worked when it was pushed. None of it ran on its own.

2:53Five decisions, and one question

Each gap came with a decision for me, five in all, each with a recommendation. How should the committee see a viewer's words? Keep a clean, shortened copy that only the committee can read. How does the committee pick? Every seat ranks the waiting tickets, and the totals go to the operations chat. I took the recommendations. Then I asked one question: does the pick include the divergence test, the check that a question will actually split the room? It did not. It was in the design and missing from the plan. Approving a plan quickly is fine only if you still know the system well enough to notice what is not in it.

3:36And a character in a rut

One more item went on the list that afternoon, and it was not a bug. Brent, the Chair, had started saying the same things every episode. A number of new subscribers. Two of everything. Viewers who watch more than one episode will hear it. The job was to vary him without changing his judgment. Hold that; it comes back at the end.

3:57The first ticket through the front door

With the helpers' work merged, I went to the website and wrote a ticket. The box was too small to edit in, and the handle to resize it was hard to see, even for me. That went on the list. The count of tickets in the bag did not move right away; it refreshes every ten minutes, and I left it, because the real acknowledgement is the receipt by email. Except the receipt never came. Every mail worker was failing to sign in. The mailbox's grant had expired, most likely when we threw the kill switch for real. A switch has a tail. Whatever it revokes has to be granted again, and you find out when someone's receipt does not arrive.

4:41The pick goes wrong, and gets undone

Then the first live pick, and it chose wrong. It drew an episode ahead of the one it depends on. Only two candidates reached the vote, and the saved record had lost the scores. My ticket had tested four to nothing: every model agreed, which makes a boring debate. Worse, the test had asked the models cold, with no character at all, and cold, most questions split nothing. The humor score failed on every ticket because each reply was cut off at its length limit. So we rolled it back. Nothing debated, nothing rendered, nothing uploaded, and the operations chat told my phone so. Undo is part of the feature. Then the fix I asked for: give each model its full character for the test, and keep the test short of a whole debate.

5:38The round trip

Second attempt. I wrote a new ticket, about a roommate who says please and thank you to a smart speaker. Good manners, or a waste of breath? The scanner cleared it in seconds and filed it as a ticket, not a comment. This time the test ran in character, and nothing came out four to nothing; the room split two to two. The committee picked my ticket for the real draw. Forty-one model calls, all of them answered, the record complete. And then the committee floor, the chat space that mirrors the debate as it happens, lit up with the four seats arguing a letter that came in through the front door.

6:17What review still caught

It still did not go straight through, and that is the point of the gate. When I watched the render, the audience never heard the letter the seats were arguing from, and the ticket number was read out digit by digit. Both got fixed and the lines re-voiced. I also approved it once by mistake, before the fix, and there was no way to take that back. There is now: an undo, within half an hour, if nothing downstream has moved. That night the episode went up.

6:52The morning after

And the next morning, nobody sent anything. The Monday draw ran on its schedule. The committee picked a seed ticket, argued it, rendered it, and posted it for review. This is what that looks like: a message in the operations chat with a link that lasts seven days, and three answers. Approve moves it on. Reject spends the slot. Fix hands the note to an agent on the largest model, which repairs the episode and posts it back to the same thread. I watch, I mention the clerk, I type approve. The rest of the pipeline adds the thumbnail and captions, uploads it, and once it is public, posts the first comment and reminds me to pin it. The episode page on the show's site publishes itself.

7:44Close: now it runs

Which brings back Brent. His lines came from his character sheet, where the habits were written as quoted examples, and a model will repeat a quoted example word for word. We rewrote them as descriptions, removed the repetitive phrases, and left his judgment alone. That is what the work is now. Not building. Tuning, reviewing, approving from a phone. The whole workflow is functioning as designed, and the show publishes on its own, every Monday, Wednesday and Friday. Done was a claim. Running is the proof.