AI coding agents can now write working animation code in CSS, Motion, GSAP, SwiftUI and other frameworks. For a hover effect, a modal transition or a simple loader, the result is often usable after a quick review.
The weaknesses show in details that make motion feel right. Agents tend to repeat the most common patterns, such as the same short ease for everything. They often handle the happy path but not interruptions, for example what happens if a user closes a sheet while it is still opening. They frequently skip reduced-motion handling and states such as loading, error and disabled.
You can close much of this gap with context. Give the agent your motion tokens, list the states each component needs, and ask explicitly for reduced-motion support. Ask it to explain its timing choices so you can check them.
Another approach is to let the agent work in a motion tool instead of writing everything in code. With an MCP server, an agent can create and edit animations in a tool that has states and physics built in. ExodeUI, which we make, has an MCP server for this, and the result is an open .exode file you can review in the editor.
Either way, keep a person in the loop for motion that users will see often. Watch it on a real device and tune it by feel.