ageation
An engine for agentic video creation: it reads a structured LaTeX source and produces short, narrated Manim videos.
ageation is a layered, file-based pipeline. Every stage reads ditections and writes files — concept maps, narrated scripts, generated scenes — so every step is inspectable, diffable, archivable, and re-runnable. Content hashes chain each layer to its source, automated gates catch drift and layout errors before render time, and a human reviews exactly two artifacts per video: the concept and the script. The machines do the rest.
Why now
Mathematics already lives in a machine-readable form, and the tools that animate it are programs. The moment agents became fluent in code, the whole path from a theorem on a page to a narrated animation on a screen became something an agent can walk — with a human deciding what is worth saying, and checking that it was said right. ageation is that path, made into infrastructure: not a one-off script, but a pipeline with provenance, review gates, and a house style that survives from one video to the next.
How it works
- Vendor the source. The upstream LaTeX notes are copied into an editable working set, each file stamped with the hash of its original — the notes themselves are never touched.
- Distill the concept. An agent reads the chapter and writes a concept map: what the video should teach, which ideas earn a beat, what is deliberately out of scope. A human reviews it (optional).
- Write the script. The concept becomes narration — beat by beat, with sync markers, a runtime budget, and a cut list. A human approves it (optional).
- Generate the scenes. The script becomes Manim code in the series' house style. A geometric lint replays every animation and flags overlapping layouts before anything is rendered.
- Build and measure. Free draft renders for iteration; one command renders, stitches, and measures the final video — and checks the measured runtime against the script's contract.
Those five stages end at a draft, not a deliverable. A human watches it and leaves notes in a Claude or Codex session: this beat runs long, that label overlaps, soften the wording here. The agent folds each note back into the script or the scenes and re-renders, and the loop continues until the video reaches deployment readiness. Only then is it rendered at full resolution and published to a video platform.
What else
Notes are only the beginning. The same pipeline ships with project shapes: a course fans a book out into a video series; an article distills a single paper — point it at an arXiv preprint and the same five stages produce one tight explainer, with the same provenance and the same review gates. If it is written in LaTeX, it can be animated.
Built on
ageation stands on excellent open tools: Manim Community renders the animations, manim-voiceover binds narration to them, LaTeX typesets the mathematics, FFmpeg stitches the videos, and uv keeps the environment reproducible. Narration is synthesized with gTTS for free drafts and the OpenAI text-to-speech API for finals.