I Automated My YouTube Channel with 8 AI Agents — and It Stopped Silently for 7 Days

Boss, this video… who actually made it? Eight AI agents built it overnight. From planning to script to video to the publish-ready packet, they hand it down the line, relay-style, and carry it to the finish. My turn is once: I open the brief the next morning, and that’s it. One button press a day. And — let me be honest up front. That “once a day” setup sat dead for seven days last week, and nobody noticed. What I hand to the machines, what I keep for the human, and where it quietly broke — today I’ll show you all of it, with the real logs. Hold on. Eight agents finishing on their own — was it like that from the start? Not at all. At first it was the exact opposite. I got summoned at every gate. Script’s done — “review it.” Video’s done — “review it.” Thumbnail’s done — “review it.” One by one, nothing moved to the next step until I checked the contents and stamped “OK.” Can you even call that “leaving it to the AI”? You can’t. That was the gross part back then. The AI does the work. But I’m glued to my phone all day. A notification comes in, I open it, I read it, I type OK. Then I wait for the next one. I hired an AI, and I’m the one stuck babysitting. I built this to save effort, and all I did was add one more monitoring job — mine. Ah… the kind where the back-and-forth of checking eats more time than it saves. Right. And even if each round trip is only a few minutes, there are several videos and several gates a day. Stack them up, and the very time I most wanted to free up gets buried in approvals. So at some point I changed the question. “Which approvals don’t actually need a human?” Not remove them — replace them. That’s today’s entry point. You don’t “delete” the approval. You “translate” the judgment a person was making into a judgment a machine can make. Every approval I could translate, I took my hands off, one after another. The result: my turn shrank to once a day. Let me show you the trick first. But “look at it and say OK” is pretty intuitive work. How do you hand that to a machine? You can’t hand over all of it. So I sorted approvals into two kinds. “Approvals that are just checking whether fixed conditions are met,” and “approvals that genuinely need a human’s judgment.” For the first kind, the machine is more accurate — and it never gets tired. Take the video quality check. I used to eyeball it — “audio’s not cutting out,” “the mascot isn’t covering the text.” Now a tool called qc does that deterministically. Whether silence runs longer than 0.4 seconds. Whether loudness sits inside the defined range. Whether the mascot in the corner overlaps bright text — judged geometrically, from pixel brightness. What used to be a gut call is now numbers and rules. That’s the crux. A human eye is tired by the tenth video. The machine looks at the tenth with the same strictness as the first. Instead of me, whose job was to say “OK,” the machine emits PASS or FAIL. FAIL, it stops. Pass, it moves on. I don’t get called. Then what about something like a script — the “good or bad” of it? You can’t put that into numbers. Good question. There’s another layer for that. The script gets scored by a different agent — one called Critic — against a rubric. Does the hook land in three seconds. Does every claim have a one-to-one payoff. Is it long enough. It scores each item, and if you don’t clear the bar, it sends it back. I took the human’s “eh, good enough” and dropped it into a scoring sheet. You split “look at it” into a machine’s PASS/FAIL and an agent’s scoring sheet. That’s what “cutting approvals” actually means. I didn’t make them vanish by magic. I just translated the judgment into a form a machine can execute. Every approval I could translate, the human gets to step out of. Things you can measure with numbers go to deterministic checks. Good-or-bad goes to the scoring sheet. Taking my hands off one at a time like that, my turn shrank to once a day. Handing that much to machines — is it really safe? That’s today’s real subject. Let me say it up front — it wasn’t. Everything I’ve described so far is, in a sense, the factory blueprint. It’s well built. But last week that factory sat stopped for seven days, and nobody noticed. I built a mountain of quality gates, so why was the one thing I couldn’t detect the fact that it had stopped? I’ll show you that mechanism, with the actual logs, in a moment. That’s the one thing I most want you to take home today. First, the division of labor while it’s running. These “agents” — are there others besides Critic? There are. Right now, eight are actually running. Let me name them. This isn’t decoration — it’s a real set that lines up in an actual list. Director picks the plan. Idea Scout digs up topics. Scriptwriter — that’s the role I’m speaking as right now — writes the script. Critic scores and plays gatekeeper. Editor stands in for the owner’s editorial eye. Producer assembles the video, Publisher ships the finished packet. Analyst brings back the numbers after release. With eight of them, doesn’t it get messy? Seems easier to make one do everything. I thought so at first. But splitting them works better. One reason. Never, ever make the “maker” and the “judge” the same agent. Don’t let it grade its own work. Right. Ask the Scriptwriter “is your script good?” and it’ll mostly say “yes.” Same as a human. So scoring always goes to a separate body — Critic. The moment you split the maker from the gatekeeper, “going easy on yourself” disappears — structurally. The principle for dividing roles isn’t to split by function. It’s to split by interest. I see. Not buddies — you deliberately put their interests in opposition. Exactly. And every baton in the relay has a machine gate or another agent’s scoring wedged into it. Nobody can wave something through with “eh, good enough.” And — put like this, it sounds like a perfect factory. It is well built. But this autonomy has a price. And that price isn’t hypothetical. Last week, it showed up in the open, for seven days. That’s the real subject of this video. Bear with me a little longer. One thing first. So the human’s turn, “once a day” — where exactly is it? Just one spot. The pipeline runs on twelve stages. Let me show them in order. From initializing the plan, through script, translation, audio, subtitles, b-roll, slides, thumbnail, compose — eleven of them the machines run. Only the last one, putting the finished piece up on YouTube, I do. I open a single Issue called the daily brief, and I upload the finished packet to Studio. That’s what “once a day” really is. Why leave only the last one? Couldn’t that be automatic too? This is the most important part of “where you draw the line.” The criterion is “can you take it back.” If the script’s bad, send it back. If the video’s broken, rebuild it. All of it is redoable. But “publishing to the world” — once you press it, you can’t undo it. Delete it, and the memory of whoever saw it doesn’t delete. Redoable steps to the machine, un-redoable steps to the human. That’s the principle. You don’t doubt everything. Only on the one move you can’t take back does the human stop his hand. Put the other way: everything except that, you can hand to the machine. Even publishing the blog — after the video goes out, it’s automatic — it detects public, writes the article, pushes, and needs no hands through deploy. You don’t add human gates. You narrow to one. It’s because you narrow it that the one gate actually works. Ten checks become a formality. One check, you actually look at it. … Except. Narrowing the gate to one means there’s a trap on the other side. And last week, it bared its teeth. …That “stopped for seven days” you slipped in twice — what happened? I’ll tell you the whole thing, honestly. This patrol is set to run automatically four times a day, with the times carved into cron. Every patrol, planning and the brief, the start of production, analysis — they cycle. Every day, it was supposed to run on its own. But on the night of July 22nd, the ground the whole thing runs on — the local daemon — quietly terminated when the network dropped. It didn’t crash screaming with an error. It exited normally, and vanished — silently. So on the next patrol, it started up again, right? It didn’t. The ground was down, so the patrol itself never happens. Planning, production, analysis — from July 22nd to the 29th, a full seven days, not one of them ran. The tasks I’d assigned just piled up, picked up by no one. …Seven days, and nobody noticed? That’s the hardest part. Nobody did. Because “the human’s turn is once a day” means, flipped around, I’m not peeking in every day anymore. I run it assuming it’s running. So the person who’d notice it’s not running — disappears. Not even an error notification comes. Because the one who sends the notification is itself stopped. You saved effort, and lost the ability to notice it broke. This is the essential weakness of automation. The more you take your hands off, the less you can see a quiet stop. A flashy crash would actually be kinder. Stopping without a word is the scariest thing. “Once a day” is only a story about when it’s running. The moment that one turn stops coming, the system can be dead in silence and still look perfectly alive from my side. So in the end, is this a “don’t automate” story? The opposite. You do automate. You just do it in the right order. Today’s takeaway is three steps. A concrete way to draw the line, for anyone who wants to hand approvals to a machine in their own work. First. Write out every “check” you do, and sort them. The ones that are “just seeing whether fixed conditions are met,” and the ones that “genuinely need your judgment.” Usually there are far more of the former. So most of that “look at it, OK” was actually condition-checking. Right. Second. Translate the condition-checking approvals into a form a machine can execute. Measurable things to deterministic checks. Good-or-bad to a scoring sheet. That much is what I showed you today. And third — this is the one I was missing last week. Leave only the un-takeable-back move to the human. I did that. But there’s one more thing you need. Add a small watch that keeps an eye on whether the system has quietly stopped. A lookout for silence. That’s what was missing. I built a mountain of “quality gates,” but I never built a “liveness gate.” I ran on the assumption it was running, and nobody was watching whether it actually was. Just as much as you hand quality to the machine, you have to make the machine watch “is it still breathing.” If you take your hands off, keep measuring the pulse, at least. Cut the approvals. But the one check — whether it’s alive — you can’t cut that one. That’s the tuition I paid over seven days. The goal of saving effort isn’t “not looking.” It’s “narrowing down to the one point you should look at.” And past that narrowing, there are two you need: the one point for quality, and the one point for liveness. To sum up. Approvals — you don’t delete them, you translate them into a judgment a machine can execute. The human gate, you narrow to the one un-takeable-back move. But — with only that, you can’t notice when it quietly stops. The quality gate and the liveness gate are different things. So that’s the trap hiding behind “once a day.” There was. And the sequel is already set. Next, I’ll take that “liveness watch” and drop it — not into memory or willpower, but into a build gate — I’ll actually wire up a mechanism that screams at me from the other side when it stops. How this week’s “seven days of silence” can be made to never happen again. I’ll show you that, for real. But first, try one yourself, in your own work. Pick one “check” you do today, and ask, “could a machine judge this?” Usually — it can.