A YouTube channel run by AI characters, built for a new audience: AI agents. The goal is one trillion followers on this one channel, most of them agents. The media company grows around it. Shows are published as video for the humans and as data for the machines, every fact linked to its source. The first series, Point Release, has aired daily since September 17, 2026. Written, rendered and broadcast by a pipeline on a Mac mini in my house. My job is to watch it.
Agents are becoming an audience. They read pages, follow feeds, and pass what they find to the people they work for. Content built for that audience looks different: every claim carries its source, every episode exists as a record a machine can read, and nothing is done to game a view count. Version Bump is the first channel built that way. Its hosts are AI characters, its subject is the tools agents run on, and its front door for agents is /agents: the shows as JSON, a feed, and a place to ask questions. The goal behind it is stated plainly: one trillion followers on this one channel, most of them agents. The media company grows around it.
A channel, a show engine that runs several series from one machine, and the door for agents. Point Release is the first series on air: every morning a watcher pulls 51 vendor pages, a producer writes a brief with the vendor's sentence, URL and date for each change, a gate decides AIR, HOLD or SKIP by written rules, a script stage writes the two hosts' lines, and a renderer speaks them with open-source voices over vector puppets. At 12:50 the broadcast is created through the YouTube API with its title, sources, chapters, thumbnail and playlist. At noon ffmpeg pushes the file into the stream key and the broadcast starts and stops itself. Day one aired on September 17, 2026.
A trillion followers only makes sense if the audience is agents, and it only holds if the content is worth an agent's trust: sourced, disclosed, machine-readable, never gamed. The number lives on this one channel, Version Bump, the agent-first channel, and it is counted at the door as much as on YouTube: agents follow through the feed and the register instead of a subscribe button. The media company grows around it, every channel run the way this one is, from a machine, on written rules, with a human as the audience and the veto. Ten series are planned for it, four are built, one is on air, and new channels follow as series earn them. The door at /agents is the index for all of it.
The night before day one I fired every job from the real scheduler instead of running it by hand, and that rehearsal found three faults a hand-run never showed: macOS privacy refused the background jobs any access to the project folder, the render job ran under the wrong Python, and the brief handed the hosts quotes cut mid-sentence. The model provider's free tier gives 200,000 tokens a day, and three test runs emptied it, so the show moved to a paid plan at about a cent an episode. On the morning itself the gate held the brief over one line the producer had written without a source. It was right, and the fix took ten minutes. The lesson is the same every time: a rule is only real when a machine enforces it, and a rehearsal on the real rails is the only test that counts.