Study of agent-driven animation: framework selection, an /animation agent skill, smoke tests and showcases in Blender, Godot, Three.js and Motion Canvas
  • Python 36.3%
  • GDScript 33.2%
  • TypeScript 27.8%
  • JavaScript 1.5%
  • Shell 1%
  • Other 0.2%
Find a file
Luciano Nery 68753c7987
Publish the agentic animation study
Framework research and selection, the /animation agent skill, toolchain
smoke tests, a four-route showcase, a 2D route comparison and the Warrior
vs Dragon study, with source, briefs, final renders and documentation.
2026-10-20 20:40:00 +00:00
docs Publish the agentic animation study 2026-10-20 20:40:00 +00:00
media Publish the agentic animation study 2026-10-20 20:40:00 +00:00
projects Publish the agentic animation study 2026-10-20 20:40:00 +00:00
skill/animation Publish the agentic animation study 2026-10-20 20:40:00 +00:00
tools Publish the agentic animation study 2026-10-20 20:40:00 +00:00
.gitignore Publish the agentic animation study 2026-10-20 20:40:00 +00:00
AGENTS.md Publish the agentic animation study 2026-10-20 20:40:00 +00:00
CHANGELOG.md Publish the agentic animation study 2026-10-20 20:40:00 +00:00
LICENSE Publish the agentic animation study 2026-10-20 20:40:00 +00:00
README.md Publish the agentic animation study 2026-10-20 20:40:00 +00:00
ROADMAP.md Publish the agentic animation study 2026-10-20 20:40:00 +00:00
SECURITY.txt Publish the agentic animation study 2026-10-20 20:40:00 +00:00

Agentic Animation Study

Which frameworks let an AI coding agent make animation on its own, and how should that agent work? This repository is the full study: the framework research, the agent skill it produced, the installs and smoke tests, and three rounds of showcase animations, each written, rendered and checked by a coding agent following that skill.

Warrior vs Dragon, Godot 2D

Warrior vs Dragon (20 s, Godot 2D). Click for the video; the Motion Canvas version is drawn from the same baked choreography.

Findings

  • Four routes cover nearly everything: Blender for 3D video, Godot for action and game feel, Three.js for 3D on the web, Motion Canvas for programmed 2D video. Blender's Grease Pencil covers a hand-drawn 2D look without a fifth tool. Why these, and why not the other ten.
  • What matters is control, not features. An agent needs plain-text scenes, a headless render, frames that come out the same every time, and single frames it can render and inspect. Tools that only export through a GUI, such as StykAnimate and Stykz, were rejected however good their output.
  • The agent only sees animation through frames it renders and opens. The skill's core step, Look, renders stills and numbered contact sheets for every beat and checks them against a timed beat sheet. In this study that step caught an off-screen ball, fire aimed at the ground, a leap through a dragon's head, wings standing up like sails, a full-screen glow haze and a victory squeezed into the last frames. The skill.
  • Bake the choreography once, draw it anywhere. For the final study, one Python file evaluates poses, IK, cloth simulation, hit-stops and the camera, then writes every frame's shapes to JSON. Godot and Motion Canvas only draw them, so both versions match frame for frame. How.
  • Render cost spans more than three orders of magnitude at 1080p on a GTX 745: Cycles ≈ 40 s per frame, Grease Pencil ≈ 2.3 s, Godot Movie Maker ≈ 0.25–0.33 s, Three.js capture ≈ 0.08 s, Revideo ≈ 0.024–0.04 s.

The showcases

Study Videos Write-up
Four routes: one scene per framework, chosen to show what each does best Blender: Collapse · Godot: Arena · Three.js: Field · Motion Canvas: Loop docs/04
2D routes: the same stick-figure fight in three 2D styles Grease Pencil · Motion Canvas · Godot 2D · side by side docs/05
Warrior vs Dragon: a 20-second fight, game look and illustration look Godot 2D · Motion Canvas docs/06

The four-route showcase: Blender, Godot, Three.js, Motion Canvas

Repository layout

Path What it holds
docs/ The write-up, in reading order, plus accepted decisions
skill/animation/ Snapshot of the /animation agent skill: workflow, route guides, methods, extension contract
projects/ Every project's source and BRIEF.md: smoke tests, the four-route showcase, the 2D study, Warrior vs Dragon, Grease Pencil API probes
media/ Final renders (H.264)
tools/grid.sh Tiles chosen frames into a labelled contact sheet

Documentation

  1. Framework selection: the candidates, what an agent needs, and the verdict on each.
  2. The /animation skill: the base workflow and how specialised skills extend it.
  3. Toolchain and smoke tests: versions, installs, and what each route's first test changed.
  4. Four-route showcase: one scene per framework, with measured render costs.
  5. 2D routes: choosing a 2D route, Grease Pencil on Blender 5.x, the shared fight.
  6. Warrior vs Dragon: rigs, baked choreography, the two looks and the fixes.
  7. Gotchas: every trap hit during the study, with the fix.
  8. Reproducing the study: set up the tools and re-render any project.

Quick start

Re-render the final study (Python 3, Node 20+, ffmpeg, Godot 4.7; full setup in docs/08):

cd projects/warrior-dragon
python3 choreo.py bake                       # writes bake.json: every frame's shapes
cp bake.json motion-canvas/src/
(cd motion-canvas && npm ci && DISABLE_TELEMETRY=true npx tsx src/render.ts --out renders/full.mp4)
(cd godot && godot --path . --write-movie renders/frames/f.png --fixed-fps 60 --quit-after 1200)

Test machine

Debian 13 (Linux 6.12), Intel i7-4790S, 15 GB RAM, NVIDIA GeForce GTX 745 (4 GB, driver 580). Every timing in these docs was measured there, often with other renders running at the same time, so treat them as relative.

License

MIT © Luciano Nery. Built by Luciano Nery.