OpenAI's devrel cut a launch video in HyperFrames with an agent that picks up 'feedback ready' notes and re-edits on its own

On Oct 22 Romain Huet posted a vertical video for the Decisions API reaching general availability and said he 'had Codex reimagine a very important app'. The post passed 280 likes within hours. Brent Schooley quoted it to explain the production. The video was edited entirely in the cloud using HyperFrames, coordinated by his agent (he calls it his 'dot') through an automatic feedback loop in OpenAI's internal review app, Cutroom. Romain reviewed each edit and marked it 'feedback ready'. The agent picked that up and kicked off a new revision.

The takeaway is the loop, not the tool names. Cutroom is internal, but the pattern is simple: render a cut, post it somewhere a human can leave timestamped notes, use a status flag as the trigger, and have the agent pull the notes, revise the HyperFrames composition and re-render. A Frame.io-style comment thread, a GitHub issue label, or a Notion status field could all be the trigger. The human only reviews.

HyperFrames shipped v0.8.186 the same day. Capture motion blur, which was opt-in yesterday, now reaches renders and Studio. On pages it can't blur safely it switches itself off and reports why, and `check` tells authors why blur is off. The release also adds a CLI feature to 'read a whole recording as voice, effect and music lines' (PR #6056), and `init --video` now frames a new project to the clip's shape. Both help when an agent starts from raw footage.

We haven't seen Brent's harness, and neither post shares code, so treat the setup as described rather than reproducible. If you build your own loop, cap the revisions per cycle and keep the human 'feedback ready' gate. Without it, the agent will keep re-rendering on its own guesses.