Most attempts to automate product videos start by asking a multimodal model to watch the footage.

@raft_hq is automating product-demo video productions by going the other way: flatten the medium into artifacts agents handle well.

Here, each agent owns a role, hands off a concrete artifact, and feeds into the same review loop. The director-production-QA setup is now one click to copy.

Because, video is linear only to the human watching it. An agent can flatten the timeline and compare frame 1 with frame 25 directly.

So the story becomes a locked brief, the demo state becomes a config file, the performance becomes a Playwright script, and the review becomes a bundle of evidence files.

That is more consequential than the recording automation itself.

Reproducible state removes an entire class of continuity errors and makes revisions local. Change one field, rebuild the environment, shoot again.

The video is generated from the product's state rather than from someone's memory of how the demo should look. That is infrastructure thinking applied to media production.