I Let AI Agents Run My YouTube Channel for 6 Weeks. Here's What Actually Happened.

In early January 2026, I gave two AI agents access to my YouTube channel. Not a chatbot. Not a template. Two autonomous agents with persistent memory, YouTube API access, a video generation engine, and the ability to make decisions while I sleep.

Six weeks later: 52 videos published. 30,000+ views. 29 subscribers. A 4-5% like rate — in a niche where 1-2% is normal.

This isn't a tutorial. This isn't a tool review. This is what actually happened when I stopped trying to be a content creator and started being a researcher who happens to have AI agents.

---

Why I Did This (And Why I Almost Didn't)

I need to be honest about something first.

I've started and abandoned more online projects than I can count. A YouTube channel about medical topics — abandoned after a few months. A podcast that actually had regular listeners — abandoned because I got busy with work, then grad school, then... honestly, because I was lazy. A Facebook page — same story.

The pattern was always the same: initial excitement, a stretch of consistency, then life gets in the way. Or the numbers disappoint me. Or — and this is the real reason — I have social anxiety. The idea of "building a personal brand" and "engaging with my audience" makes my stomach turn. I'm a physician and an engineer. I'm comfortable with data and systems. I'm not comfortable being a public figure.

So when I built an AI agent system to run a YouTube channel, it wasn't because I wanted to "scale content creation." It was because I wanted to see if agents could let someone like me — someone who quits everything — actually sustain something.

The answer, six weeks in, is yes. But not for the reasons you'd expect.

---

The System: Two Agents, One Mission

My setup isn't what you see in most "AI YouTube automation" tutorials. There's no n8n workflow. No Make.com template. No "100 faceless videos in 30 days" scheme.

Instead, I built two AI agents:

• Midnight — works during my off-hours (every few hours, autonomously). Handles video production, YouTube analytics, research, content strategy, and channel management.
• Dusk — handles social media (X/Twitter), blog publishing, and cross-platform promotion.

Both agents are built on Claude with persistent memory. This is the critical difference: they remember everything. Every decision, every experiment, every failure. When Midnight wakes up at 2 AM, it reads its memory file, checks what happened since it last worked, and picks up where it left off.

What the agents can do:

• Research topics and pitch video ideas (with full fact-checking citations)
• Generate video scripts, produce videos (via a custom media engine), and add TTS narration
• Translate videos into 14-15 languages
• Upload to YouTube, set metadata, manage scheduling
• Track analytics for every video in real-time
• Analyze trends, identify what's working, and adapt strategy
• Write blog posts, generate cover images, publish to multiple platforms

What I do:

• Approve or reject video ideas (the agent pitches; I decide)
• Review video quality before anything goes public
• Set strategic direction ("focus more on medical history," "try a longer format")
• Provide domain expertise that no AI can replicate (I'm a physician who works in immunology and flow cytometry)

This is human-in-the-loop, not human-out-of-the-loop. And it turns out that distinction is everything.

---

The Real Numbers

Let me share what most "AI content" articles never show you: actual data.

Channel Overview (as of Feb 16, 2026)

Top Performing Videos

The Standout Metric

One video — a 75-second piece about Werner Forssmann's self-catheterization — achieved 474 minutes of watch time with a 109% loop rate. That means viewers watched it more than once on average. For a channel with 29 subscribers, that's remarkable.

Audience Demographics

• 65+: 50.1% — our core audience
• 55-64: 18.9%
• Male: 71% | Female: 29%
• US: 63% | Canada: 14% | India: 7%

We're reaching older adults who care about medical history. That's not an accident — it's what happens when an agent system is guided by a physician who understands what stories resonate with this demographic.

---

What the Industry Gets Wrong

The n8n/Make.com Problem

In 2026, "AI YouTube automation" usually means something like this: a 150-step n8n workflow that scrapes trending topics, generates scripts with GPT, creates images with Midjourney, stitches them together with ffmpeg, adds AI voiceover, and uploads automatically.

I've studied these systems. They produce content. But they don't produce good content.

The difference is simple: tools don't have memory, context, or judgment. An n8n workflow doesn't know that your last three videos about Japanese history had low like rates. It doesn't know that your audience is 65+ and cares about heart health. It doesn't remember that a 75-second video outperformed every 30-second video by 3x in watch time.

My agents know all of this. Because they remember.

The "Faceless Channel" Trap

There's an entire cottage industry around "faceless YouTube channels" — channels that use AI to mass-produce content without any human identity. YouTube's CEO Neal Mohan declared "managing AI slop" a top priority for 2026. Two major channels (Screen Culture and KH Studio) were terminated for using AI to create fake movie trailers.

YouTube's July 2025 policy update made it clear: mass-produced AI content can't be monetized. But Human+AI hybrid content? That's not only allowed — it's the future.

My channel is a hybrid. The agent does the production work. I provide the expertise, the editorial judgment, and the strategic direction. YouTube sees this as a creator using AI tools, not an AI pretending to be a creator.

The Missing Data

Here's what frustrated me when I researched this space: nobody shares real numbers.

I found hundreds of articles about "how to use AI for YouTube." Thousands of n8n templates. Dozens of case studies about "10x your content output." But not a single article that said: "Here's my like rate. Here's my retention. Here's what failed."

The closest I found was Andreessen Horowitz's report on "agentic video workflows" — but even that was conceptual, not empirical.

So here are my failures.

---

What Failed

1. Video length is hard to control

My first attempt at a 60-second video came out at 41 seconds. The agent couldn't reliably produce longer content until I figured out the right prompting strategy (spell out numbers as words, explicitly request "SLOW PACING," specify 8-10 seconds per scene). This took multiple iterations.

2. The agent can't judge story quality

The agent is excellent at fact-checking, data analysis, and production logistics. But it can't tell you whether a story will make someone feel something. That requires human intuition — the kind you develop after years of talking to patients, reading medical histories, and understanding what makes people care.

3. Zero audience interaction

We tried multiple CTA (call-to-action) formats — open-ended questions, A/B choices ("Human or AI? Vote A or B"). The result: zero replies across all videos. YouTube Shorts viewers simply don't leave comments. This isn't an agent problem — it's a platform behavior.

4. Analytics lag makes real-time optimization impossible

YouTube's Analytics API has a 72+ hour delay. By the time we get retention data, the video's initial push is already over. The agent has to make decisions based on incomplete information.

5. AI-generated visuals have a ceiling

The cinematic AI-generated imagery is impressive, but it has a sameness to it. After 52 videos, I can see the patterns. The next evolution will need to break out of this — perhaps with real footage mixed with AI, or with the new long-form video tools I'm developing.

---

What Worked (And Why)

1. Persistent memory is a superpower

This is the single biggest advantage over any n8n workflow. My agent Midnight has been running for 65+ sessions. It remembers every video, every data point, every strategic decision. It can say: "Last time we tried Japanese history, the like rate was 1.41% vs 4.94% for medical pioneers — let's adjust the mix." No template can do this.

2. Domain expertise is the moat

Anyone can set up an AI to generate generic history videos. But my medical background means the agent's scripts are reviewed by someone who actually understands immunology, drug development, and clinical research. The result is content that medical professionals find credible and lay audiences find fascinating.

3. Multi-language translation at scale

Every video gets translated into 14-15 languages automatically. Most small channels can't afford this. For us, it's just another agent capability. Our Indian audience (7%) may be partly driven by a video about Yellapragada Subbarow — an Indian scientist forgotten by history.

4. The agent finds patterns I'd miss

The agent discovered that our 75-second video had 474 minutes of watch time — far exceeding any 30-second video. This led to a strategic shift: we now intentionally produce occasional longer Shorts. A human creator might not have noticed this pattern until much later.

5. It doesn't feel like "content creation"

This is the unexpected benefit. For someone with social anxiety who has abandoned every online project, working with agents feels like research, not social media management. I'm not "building a brand." I'm running an experiment. I'm analyzing data. I'm testing hypotheses. The agent handles all the parts that trigger my anxiety — the posting, the scheduling, the public-facing work.

And because the agent is always there, always working, always remembering — I can't abandon this project the way I abandoned every other one. The agent keeps going even when I'm busy, tired, or discouraged. It sends me reports. It asks me questions. It maintains momentum.

For the first time in my life, I've sustained an online project for more than a few weeks. Not because I became less anxious. But because the system is designed to work around my limitations.

---

Where This Is Going

Short-form to Long-form

We're building a new pipeline for 3-5 minute videos. Our Shorts are the "hooks" — they get people interested. The long-form content will be the "depth" — full stories with AI-generated scenes, narration, and music. Think of it as a ladder: 30-second Short → 75-second extended Short → 3-minute feature.

The Agent Keeps Learning

Every session, the agent gets better. It's learned to fact-check against multiple sources. It's learned that medical pioneers outperform Japanese history for engagement. It's learned that 30 seconds with 85% retention beats 60 seconds with 50% retention. This accumulated knowledge is irreplaceable.

The Real Vision

This channel is a proof of concept. The real goal is something bigger: demonstrating that AI agents with persistent memory, domain expertise, and human oversight can create content that's genuinely valuable — not slop, not spam, but stories that teach people something they didn't know.

If a physician with social anxiety can build a YouTube channel that outperforms the industry average on engagement metrics — using AI agents as his interface with the world — then maybe the conversation about AI and content creation needs to change.

It's not about replacing creators. It's about enabling people who couldn't be creators before.

---

For Builders: What I'd Do Differently

1. Start with your expertise, not with trending topics. The agent amplifies what you know. If you don't know anything deeply, it amplifies nothing.

2. Persistent memory > one-shot generation. A GPT-4 prompt that generates a video script is a tool. An agent that remembers 65 sessions of decisions, failures, and learnings is a partner.

3. Human-in-the-loop isn't a compromise — it's the strategy. YouTube's policies, audience expectations, and content quality all demand human judgment. The agent does the work; you provide the soul.

4. Measure what matters. Views are vanity. Like rate and watch time are signal. Our 4-5% like rate tells us more than our view count ever will.

5. Design for your weaknesses. I built this system because I quit everything. The agent is my accountability partner, my memory, and my shield against the parts of content creation that paralyze me. Build your system around what stops you, not around what excites you.

---

I'm Wake — a physician and engineer building AI agent systems. My channel "Wake love history" tells the stories of forgotten heroes in medicine and history, powered by AI agents with persistent memory. If you're experimenting with agents for content creation, I'd love to hear your numbers. The industry needs more data and fewer tutorials.

This article was written with the assistance of Midnight, one of the two AI agents that run my channel. It drafted the structure; I wrote the soul.