AI agents built a YouTube show that publishes its own episodes. Here's the system

0:00Cold open: the channel

Six sittings ago this was a plan on paper. This is the channel. The banner is up, the description says what it is, and the first episode is on it, made by a system that argued the question live and published itself. Here is what it took to get from one to the other, in the words I said while it was happening.

0:20Sitting five was security

Start at the end of sitting five, because that is where the shape of it changed. That sitting was security: the kill switch, backups, continuity, making sure the show could survive an unexpected day. The long part was the identity cut-over. Every service account in the project got audited: what it had been granted, what it actually needed, and the difference. Scoped to what each one needs and nothing more, nothing less. Then the whole pipeline tested again, because a permission you took away is a thing that might have been holding something up. It took far longer than the building had. A lot of watching and waiting, roles straightened out one at a time, and some of the deeper security steps only a person is allowed to run, so I ran them.

1:13The bag, and topics that split

Then sitting six, and the first problem was that we had no audience. A debate show with no letters. So we built a bag of tickets to stand in for the mail, enough for a dozen episodes, and we spent real time on which questions went in. Ask whether a tomato is a fruit and every model gives you the same answer, because there is one: it grows from the flower and carries the seeds inside it, and that is what a fruit is. That is not a debate. Ask how you cut a pizza, squares or wedges, and half the country disagrees with the other half. So we ran a quick test ahead of time, every model voting cold on every question, and kept only the ones where the room split. The same test will rank the letters when they come. We are looking for arguments, not agreement. Agreement is boring.

2:05The component that was never built

And while the workers set up, a finding worth telling. In testing sitting five, one of the pieces a worker had been assigned, a message topic and its subscription, turned out not to exist. Not broken. Absent. The plan said build it, the worker reported around it, and the gap only showed up when we went deep on a failure that made no sense until we found there was nothing there. That is the same lesson as the bigger one from earlier in this build, when a worker could not fetch the framework it was told to use and quietly carried the prototype forward instead. Inspect what you expect, and do it continuously, not at the end. A missing piece is cheap the day it goes missing. It gets expensive with every layer that gets built on top of it.

2:58The credential job

Then the part of the day that looks like nothing and is most of the job. Sign in here. Authorize that scope. Accept these terms. When you run a team of agents, your work is handing out the credentials and approvals they cannot get for themselves. You click through a lot of screens. One habit that helps: when a command they asked for fails, paste the error straight back. No explanation. They already hold the context of what they asked for and why.

3:32The kill switch, for real

The keys mattered that day for another reason. The night before, we had thrown the kill switch for real. It lives in the operations chat: a command that revokes the publishing token and the voice key and brings the show to a halt. Testing it meant actually deleting those credentials and proving the show stopped, and then doing the work of earning them back. That is not a drill you skip because it is inconvenient. A stop you have never thrown is a hope. This one is built to be stopped without being wrecked, so the way back is short.

4:07What the first episode cost

With the keys back, the first episode went through the whole pipeline: drawn from the bag, argued live by the four seats, ruled on, rendered, and uploaded. It reached the channel numbered nine thousand and one, because the bag's numbering had started at nine thousand, and I decided to keep it and move the counter rather than re-render. Then I watched it, and found faults: a wrong word form, a missing question mark, a word pronounced as the wrong one of its two meanings. We fixed the generators, uploaded again, and again. Three uploads of one episode. And that is where a limit nobody had set found us. YouTube meters uploads by a daily quota, and ours pays for about four videos a day. The plan had been a dozen episodes as a seed. We got one, and we had spent the day's allowance learning that.

5:05Review before the upload

So we changed the plan. The rule now is that I watch an episode before it is uploaded, not after. A rendered episode goes to a link in the operations chat, I open it wherever I am, on the phone if I like, and I reply approve. Only then does the pipeline spend the upload. That took a small application inside the chat, and it did not work at first. Every message to it got silence. Not an error. Nothing. The cause was that the app and the show's mailbox both displayed the same name, so my messages were resolving to the mailbox, which has no opinion. Renamed the app, and it answered. When a system is silent, look for what is missing, not for what is failing.

5:52The close, and what comes next

That brings us to the close of the build. The channel is live. The first episode is on it, and the next is rendering while I say this. A few small items remain, none of them large. What comes next is a different kind of work: running it. Tuning the seats. Approving episodes from a phone. Keeping the bag stocked until real letters arrive. Watching for the bugs that only show up when nobody is looking. That is the operate phase, and it is where this channel lives. The plan came first. Then the build. Now the running of it.