Coding agents can build high-quality 3D scenes, but when asked to animate them over a long horizon, they default to scripting the whole motion in one pass, yielding unrealistic animations with objects that interpenetrate, float, and move unnaturally.
DyCraft breaks the animation workflow into steps that coding agents can reliably solve. It plans and builds the scene, then storyboards key animation frames as static 3D scenes that the agent can verify before authoring the motion between them.
On 24 long-horizon animation tasks, it produces substantially fewer physical errors than a native GPT-6 agent and prior agentic baselines, and is preferred by human raters in over 70% of pairwise comparisons.
