Trends · 9 minute read · June 2026

What AI builders are actually watching right now

I mapped 21 YouTube channels that reach engineers building AI agents in production, then looked at the videos that beat their own average. Eight clear themes fell out. Here they are, with the receipts.

Lera Leonteva
Lera Leonteva
CEO & Founder, Leo AI
Part 1 of 3
21
channels that reach AI builders
41
videos beating their own average
8
topics winning right now
48.6x
the single biggest over-performer
A 3-part series
01 · The trends 02 · The method 03 · What surprised me

I run a company that turns footage people already have into video that performs, so I spend an odd amount of time on one question: what does a specific audience actually watch, and why does one video take off while a near-identical one sinks.

Last week I pointed that question at my own audience, the engineers and technical leaders building AI agents and LLM apps in production. Not the most subscribers. Not the most polished. I went looking for the videos that outran their own baseline, because that is the honest signal. Eight themes came out clearly.

One note on the measure before we start. Throughout, a video counts as a winner when it beat its own channel's 90-day median by at least 2x. That keeps the comparison fair, so a small specialist channel and a giant like IBM Technology can sit in the same list without the giant drowning everyone out.

The eight topics winning right now

6.2xTop video

1. Agent harnesses, the new coding paradigm

The loudest theme is the harness, the scaffolding around a coding agent: its memory, its orchestration, the way it ships code without falling over. Cole Medin built almost a sub-genre out of it. His full guide to building coding harnesses did 65K views, 6.2 times his median. LangChain's open-harness take hit 7.8x. A year ago almost nobody used the word. Now it is a hook.

@colemedin · 208K subs
13.4xTop video

2. Claude Code as a build platform

Claude Code stopped being a tool and turned into a substrate. Cole Medin's video on building self-evolving Claude Code memory did 140K views, 13.4 times his median. Theo, at 540K subscribers and not an easy crowd, pulled 196K views on one question: how does it actually work. When five unrelated channels build on the same tool, it has become a platform.

@t3dotgg · 540K subs
31xTop video

3. How to build agents, from skills to lifecycle

The biggest reach in the entire set goes to the most basic question. IBM Technology's "The 7 Skills You Need to Build AI Agents" did 406K views, 31 times its median. Its explainer on what agent skills even are did 264K. The fundamentals out-traveled every clever advanced build. That is humbling, and useful.

@IBMTechnology · 1.71M subs
7.3xTop video

4. Reliability, evals, and why pilots fail

The most on-target cluster for anyone actually shipping. View counts are smaller because the audience is narrower and more senior. AssemblyAI on fixing edge cases did 7.3x. Data Science Dojo on why AI pilots fail did 6x. Databricks on closing the reliability gap did 5.5x. Fewer views, higher intent. These are the ones a team lead sends to the group chat.

@AssemblyAI · 182K subs
12.7xTop video

5. Local and open-source inference

Builders want to run things themselves. IBM's explainer on llama.cpp, the local inference engine, did 150K views, 11.5x. Mervin Praison running an agent locally with no cloud did 12.7x. This is about cost and control, not novelty.

@mervinpraison · 80.9K subs
22.5xTop video

6. Autonomous data and analytics agents

Agents that own data work are breaking out inside the enterprise channels. Databricks' end-to-end demo of an autonomous data agent did 29K views, 22.5 times its median, one of the highest jumps in the study. The promise is a dashboard from a prompt, and the audience clearly wants to believe it.

@Databricks · 159K subs
48.6xTop video

7. Voice and speech agents

Here is the one that surprised me most. A single AssemblyAI demo of streaming speaker labels did 37K views at 48.6 times its median, the highest multiple in the entire study. It is concentrated in one channel today, which is exactly what an emerging signal looks like right before everyone piles in.

@AssemblyAI · 182K subs
15.8xTop video

8. A new word to watch: OpenClaw

Vocabulary is forming in real time. IBM's explainer, "What is OpenClaw," did 206K views at 15.8x. A scrappy tutorial channel built on it too. I do not know yet whether the term sticks. But when a 1.7M-subscriber channel and a small one both reach for the same new word, that is worth tracking.

@IBMTechnology · 1.71M subs

How far the winners beat their own average

Each bar is one video's views divided by its channel's 90-day median. Higher means it outran its own baseline by more.

Voice demo 48.6x "7 skills" explainer 31x Autonomous data agent 22.5x "OpenClaw" explainer 15.8x Agent lifecycle 15.8x Claude Code memory 13.4x Local agent 12.7x Local inference engine 11.5x

Source: 90-day public YouTube snapshot, June 8, 2026.

Why this lands

The numbers do the persuading.

Builders do not trust marketing copy, they trust evidence. The story is not just the multiples. Builder topics travel to hundreds of thousands of people when they are the right topics, taught plainly.

406K
views on one skills explainer
196K
on how Claude Code works
150K
on a local inference engine

What the pattern tells me

  • Fundamentals beat cleverness. The single biggest reach was a list of the skills you need to build an agent. People are still arriving, and they want the map before the territory.
  • Proof beats polish. The reliability, evals, and local-inference videos do not have the prettiest thumbnails. They have the highest intent.
  • Vocabulary is a leading indicator. Harness and OpenClaw both went from nowhere to hooks. New words are where attention moves before the dashboards catch up.
  • Emerging beats saturated. The 48.6x voice multiple is not a fluke, it is a lane that is not full yet.

What I would do with this

If you make anything for this audience, start with the topic you assume is too obvious to bother with, because the data says obvious is exactly what travels. Show real production pain and the fix, not a happy-path demo. Use the words your audience is just starting to use. And get into the emerging lanes, voice in particular, before they fill up.

If you want the method I used to build this map, so you can run it on your own audience, I wrote it up step by step in the next post. And the honest story of doing it, including the surprises and the leads I could not close, is in the third post.

A note on the numbers. These come from a 90-day snapshot of public YouTube data as of June 8, 2026. Over-performance means a video beat its own channel's 90-day median by at least 2x, which lets a 947-view tutorial and a 400K-view explainer sit in the same list honestly. I looked at 21 channels and the 41 videos that cleared that bar. View counts are rounded.

Keep reading

How to find what your technical audience actually watches

The repeatable method behind this map, in seven steps you can run on your own audience this afternoon.

I mapped my audience's YouTube diet. Here is what surprised me

The honest behind-the-scenes, including the four findings I did not expect and the two leads I could not close.

More from the blog

Our full library of writing on video, audiences, and what actually performs.
Lera Leonteva

Lera Leonteva

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.