Giving Astra control of a video editor: An experiment in autonomous visual storytelling

I've been messing around with Astra, trying to see how well it works for editing and storytelling. I had a folder with a bunch of video clips from my travels that was just lying around, and the idea was: can I give this to Astra and have it make something fun and interesting from it? I built the API first and then basically told it to go wild, invent a story, and use the clips as source material. It's still pretty rough (extremely rough), but nonetheless, I think it's a pretty fun and interesting result. Technical description Base: A fork of Shotcut with a native C++ "inspection" API patched in. API: Line-delimited JSON-RPC over a Unix socket — ~89 discoverable editing/query operations (not all of Shotcut). Optimistic-concurrency fencing, native undo, durable mutation receipts. Brain: An LLM director ( GPT-6-astra , via the Codex app-server) that drives the editor through a bubblewrap-sandboxed Python bridge — it can only call the registered methods, never arbitrary code. Sees: Qt window screenshots + real frames/audio pulled via a separate MLT process. Hears: A local Audio Flamingo Next worker (describe / transcribe / sounds / music). Speaks: Orpheus / Qwen TTS engines, run by a separate narration runner. Image generation: ChatGPT Images 2.5 via a Codex skill. Here are the first two cuts. The first cut took about 1.5 hours, but this included indexing the shot list and other setup operations. The second cut took about half an hour. youtube.com/playlist I think it demonstrates the model's current ability to reason across many modalities, and I also think it demonstrates its current limitations when it comes to coherence. Anyone who has worked with coding agents has probably done their fair share of cat-herding, and I think that's definitely still on display here. The difference between cut 1 and cut 2 also shows just how much improvement you can get with a little nudging. All I did was watch the first cut and tell it what I thought: "This seems a bit incoherent?" "What are the characters' motivations?" As a reflection, I think for anyone calling Astra AGI, this should be ample evidence that that's definitely a milestone we have yet to cross. Astra may be extremely competent within narrow domains, but creative choices aside, it still struggles significantly with something as simple as telling a coherent story. At its current stage, I do see it working as a semi-competent editor if you give it clear direction. It could definitely speed up the production of a first rough cut. The repo is currently in a pretty rough state, but I'm working towards a checkpoint and hoping to release it during the weekend. Let me know if you are interested in forking or testing it out.