Autonomous Video Studios
HomeFront matter
EXECUTIVE SUMMARY

Two video studios that run themselves

In 33 days, from 12 October to 14 November 2026, the two systems described here made 2,959 finished videos and published them as 4,621 posts. The original programme's channels have been watched 4,434,425 times — more than twice the 2.15 million of the first edition, eleven days earlier. Nobody films, edits, writes or uploads anything: an AI agent writes each episode — as code, or since 10 November as a photoreal Blender scene — a GPU renders it, a voice model narrates it, captions are timed to every spoken word, a battery of quality gates checks it, and the system posts it on a human-looking schedule to each channel's own accounts — then reads the audience numbers back and learns from them.

2,015
episodes made by the original programme
4,421
posts on YouTube, Facebook, Instagram, LinkedIn, TikTok
4.43M
views across the original's 16 live channels
28
long-form films made
944
episodes made on ChannelForge
2,735
ChannelForge jobs run
166
ChannelForge features switched on, of 198
117
Blender episodes since 10 November

The original programme#

Ten projects and 65 channels in science, engineering, nature, sport, craft, history and the human body. 16 channels are live on YouTube, Facebook and Instagram, with a curated LinkedIn queue and TikTok (X is connected and waits for API credits); the other 49 generate every day towards launch. Each live channel posts two to four times a day — two of its slots now Blender episodes — a long-form film goes out every day, and a self-healing layer watches every slot, every GPU and every platform, making up anything that was missed and putting a button beside every failure it cannot fix alone.

ChannelForge#

The same engine as a product: anyone can describe a niche, receive channel proposals, have a brand and a renderer built, and let an autopilot make and publish episodes to nine platforms, with Mastodon and Telegram built as the tenth and eleventh. It is multi-tenant (workspaces, projects, teams, roles, approvals), metered (credits charged at cost times a plan markup), billed (Stripe and PayPal built in, switched on at launch), and open (REST API, signed webhooks, an MCP server for AI agents). 166 of its 198 features are switched on — 36 for every customer, from a hook lab and language twins to retention learning, and the rest, most of them from three studies of leading products, in the owner's own workspaces first. It makes Blender episodes too, and keeps a ledger of every failed post with a lever beside each.

How it is built#

Python control planes, TypeScript/Next.js web applications, PostgreSQL, Redis, headless Chrome with WebGL2 and Blender 4.5 for rendering, ffmpeg, Claude Fable 5.2 and 5.1, Opus 5.5 and Sonnet 5.5 through Amazon Bedrock, Google Chirp 3 HD voices, a speech recogniser on every GPU that times each spoken word, NVIDIA L4 GPUs on AWS in two accounts and two regions, Google Cloud for hosting and the media archive, Cloudflare at the edge. Everything is in Git, backed up to GitHub every four hours with a secret scanner in front, and recoverable from a written turn-key runbook.

Economics#

A short episode costs about $10.82 in model time on the original programme on average — about $10 while its agent was Claude Opus 5.5, about $16 since it moved to the stronger Claude Fable 5.2 on 10 November — and $7.02 on ChannelForge, whose standard agent is Claude Sonnet 5.5, plus one to two dollars of GPU time. ChannelForge sells credits from $49 a month and charges every job at its cost times a plan markup of 1.8 to 3, so every episode a customer makes earns more than it costs.

Key facts#

Built12 October – 14 November 2026: 33 days from the first episode to this edition
Original programme10 projects, 65 channels; 16 live on YouTube, Facebook and Instagram, with LinkedIn and TikTok
ChannelForgea multi-tenant SaaS: 9 platforms open and 2 more built, 166 of 198 features on, plans from $49 a month, a REST API and an MCP server
Made so far2,015 (117 of them in Blender) + 944 short episodes and 28 + 3 long-form films (original + ChannelForge)
Published4,421 posts by the original programme, 200 by ChannelForge's workspaces
Audience4,434,425 views and 13,481 followers on the original's live channels
AI modelsClaude Fable 5.2 and 5.1, Opus 5.5, Opus 5, Opus 4.8, Sonnet 5.5 on Amazon Bedrock; Google Chirp 3 HD voices
GPUs30 NVIDIA L4 boxes for the original in 13 pools, plus ChannelForge's own; two AWS accounts, two regions, on only while there is work
HostingGoogle Cloud (Belgium): two containers, an Arm VM, Cloud SQL, 730 GB of media in Cloud Storage, copied hourly to a second bucket
Codeabout 1.14 million lines in three trees, 1,106 test files, 1,527 commits, backed up to GitHub every four hours
Recoverya turn-key runbook (DR.md), a nightly snapshot of everything deployed, a posting guard
Statusboth systems live; ChannelForge's selling switched off until its launch