The most-viewed post in my watchlist overnight wasn't a new video model. It was a list of 14 free repos that turn a video into a codebase.

Look at how the pipeline breaks down, and you'll notice there's no single "make me a video" button anywhere in it:

• Remotion composes and renders video in React, and Remotion Skills gives your agent official guidance on building with it
• Blender handles models, rigs and real 3D renders, and Blender MCP lets Claude control the scene
• GSAP does UI motion and kinetic type on a timeline
• Three.js renders 3D in the browser on your GPU
• Chatterbox generates open-source voiceovers from plain text
• faster-whisper turns any voiceover into captions
• ComfyUI runs the image and video generation pipeline
• yt-dlp pulls reference clips and audio from the web
• FFmpeg encodes, inspects and ships the final MP4

And the piece I find most interesting: Chrome DevTools MCP, so the agent can see its own render and fix it before export.

That last one changes the game. Every other tool on the list is a pair of hands. That one is a pair of eyes. An agent that can look at what it made, notice it's wrong and go back is the difference between a generator and an editor.

This is why "AI video looks cheap" is becoming a workflow problem, not a model problem. When voice, captions, 3D, motion and the final encode are all just code, every step can be rerun, versioned and checked. You stop rolling the dice on one giant prompt and start building a pipeline you can actually debug.

The editor, the voice actor and the 3D artist aren't gone. They've just been turned into steps an agent can call.