What if the future of AI video generation isn't actually video generation?
What if, instead of asking a diffusion model to hallucinate thousands of pixels across hundreds of frames, we simply let an AI agent build a real 3D scene and render it?
Well, Mirage just demonstrated exactly that.
Meet Tesseract, a creative engine designed specifically for AI agents like Claude Code, Codex, Cursor, and ChatGPT.
And this might be one of the most interesting developments in AI-assisted filmmaking.
Mirage, the company behind Captions, has built something fundamentally different from traditional text-to-video generators.
Tesseract doesn't generate a finished video by predicting what every frame should look like. It gives an AI agent the tools to construct an editable scene, control its components, and render the result using a native graphics engine.
Here's how Tesseract works.
The system combines two main components:
Tesseract Runtime: A native creative engine accessed through a command-line tool called tsrct.
Agent Skills: Specialized instructions that teach AI agents how to create videos, motion graphics, 3D scenes, and visual designs.
The agent doesn't need to click through hundreds of menus in Blender, Unreal Engine, or After Effects.
Instead, you describe what you want in natural language.
For example:
"Create a cinematic nighttime kitchen. Rain is falling outside, the floor is reflective, and the camera slowly moves toward the window."
Your agent translates that brief into a structured Tesseract project.
And here's the critical difference: the project remains editable.
Inside the .tsrct document, the agent can control:
3D geometry: Objects, meshes, shapes, and their spatial relationships.
Materials: Physically based surfaces, roughness, reflections, and transparency.
Lighting: Individual lights, shadows, and atmospheric effects.
Cameras: Position, focal length, movement, and animation.
Timeline: Keyframes, animation curves, layers, and transitions.
Audio: Sound effects, dialogue, music, and synchronization.
Version 0.5.0 introduced editable 3D scenes, polygon modeling, physically based materials, camera animation, and path-traced rendering.
The engine supports both:
Raster rendering: Faster previews for checking composition, camera movement, and animation.
GPU path tracing: Higher-quality offline rendering with realistic reflections, indirect illumination, refraction, and shadows.
And here's where the story gets particularly interesting.
Now we're giving AI agents the ability to understand and manipulate the scenes behind those pixels. The difference between generating an image of a world and actually constructing that world.
Tesseract by Mirage: https://mirage.app/tesseract
Official GitHub repository and Agent Skills: https://github.com/mirage-hq/Tesseract
3D and path tracing release: https://mirage.app/tesseract/updates/3d-and-path-tracing
Original kitchen demonstration by Mirage: https://x.com/trymirage/status/2118749780478177788
Adobe project conversion compatibility: https://dev.mirage.app/tesseract/converters/compatibility
#Tesseract #ClaudeCode #MirageAI #AI3D #AIFilmmaking #PathTracing #AIAgents #GenerativeAI
YouTube / ABVX
