Projects, scenes, compositions, and renders: a video editor's data model
How Agent Studio structures a video project internally, from top-level project down to individual render output.
By Agent Software
A single generated clip isn't a project
Most AI video tools optimize for a single output: describe a clip, get a clip. That's a reasonable shape for a 15-second social post. It's the wrong shape for anything with more than one scene, any reuse of assets, or any review step before something ships — which describes almost all real marketing and product video work.
The core entities
Agent Studio's data model separates a few concepts that a simpler tool would collapse into one:
- Projects — the top-level container for a piece of work
- Scenes — individual segments of a project, each with its own timing and content
- Compositions — the actual timeline structure inside a scene, built on one of two composition engines
- Media — video, image, and asset files shared across scenes, not duplicated per scene
- Music — audio tracks and stems, also shared rather than scene-locked
- Renders — the rendered output of a composition, versioned rather than overwritten
- Review — a step renders pass through before they're considered final
Why the separation matters
Keeping media and music as shared resources instead of scene-local copies means a logo, a product screen recording, or a music bed can be reused across a dozen scenes without a dozen copies drifting out of sync. Keeping renders separate from compositions means re-rendering a scene doesn't destroy the previous version — it produces a new one, which is what a review step actually needs to compare against.
Two composition engines, one data model
Studio supports two different composition engines under this same project structure, so the choice of engine is a detail of how a composition gets built, not a different kind of project altogether. That keeps the mental model consistent regardless of which engine ends up rendering a given scene.
An honest status note
Agent Studio is an internal project in development, currently used to produce the suite's own launch material — it's not a public, self-serve product yet. What's described here is the actual data model behind that internal tool, not a preview of a feature that doesn't exist.
