Behind the scenes · 7 minute read · June 2026
I gave myself a day to find out what the people building AI agents actually watch, and what is pulling away from the pack. I expected to confirm a few hunches. A handful of things genuinely surprised me. Here is the honest version, including the parts I could not finish.
I looked at 21 channels that reach this audience, kept the 10 still posting, and flagged the 41 videos that beat their own channel's median by at least 2x.
That last bit matters. I was not ranking by raw views, I was asking which videos outran their own baseline. It lets a 947-view tutorial and a 400K-view explainer sit in the same list honestly. If you want the full method, it is written up here. This post is what I did not see coming.
I assumed the clever advanced builds would dominate. They did not. The highest-reach video in the whole study was IBM's "The 7 Skills You Need to Build AI Agents," at 406K views and 31x its median. The fundamentals beat every flashy thing. The audience is still arriving, and they want the map before the territory.
Harness, as in the scaffolding around a coding agent. Cole Medin has three over-performers with it in the title. Then a second new term showed up: OpenClaw. IBM explained it to 206K viewers at 15.8x. Watching a piece of vocabulary form in real time, across a giant channel and a scrappy one at once, was the most interesting hour of the day.
The biggest multiple in the entire study was not agents or Claude Code. It was a single AssemblyAI demo of streaming speaker labels: 37K views, 48.6 times their median. It is concentrated in one channel right now, which is exactly what a lane looks like just before it gets crowded.
I kept seeing Claude Code, and not from one fan. Cole Medin built self-evolving memory on it at 140K views. Theo, who is not easily impressed, did 196K on how it actually works. Five unrelated creators building on the same tool is the tell. It is not a feature anymore, it is a surface other people build on.
The number I keep thinking about
How far one voice demo beat its channel's own average. The highest jump in the whole study, and a lane that is not full yet.
Honest open items
Being honest, the day ended with loose ends. Two channels I had good reason to include, an MLOps community channel and a coding-agent creator, did not resolve cleanly from their handles, so I could not confirm their numbers. I left them flagged rather than guess and quietly fold a made-up figure into the chart. I would rather show you the gaps than paper over them.
The reason I trust the rest is a boring one. I had every channel double-checked and cross-referenced against public engineering lists and each channel's own page, and I threw out anything I could not corroborate. Including the look-alikes that seemed on-target and were not, which is a story in the method post.
And one rule about the work itself: research that only ever adds and never cuts is not research. The reject pile and the flagged-but-unconfirmed pile are where the honesty lives. If you want the clean output of all this, the eight winning topics with their numbers are in part one.
I am the founder and CEO of Leo AI. I spend most of my time on one question: how do specific audiences find and watch video, and what makes some of it travel. I write about what the data shows. More about Leo.