Planning & narrative
Decomposing a communication goal into a coherent storyboard — what to show, in what order, and why. The hard part is not generating media; it is deciding what is worth making.
Upcoming project
The Media Agent
Give it a goal — a story to tell, a concept to explain, an argument to make — and it plans, produces, and edits complete media: scripts, images, video, sound, and layout, working as one system rather than a chain of disconnected tools.
In early research · Working previews from 2027
What we are building toward
Not a chain of tools.
One system that makes things.
Today, making media with AI means assembling separate generators and hoping the pieces fit. We are studying what it takes for a single agent to hold a goal, plan the work, produce each element, and edit the whole into something coherent — the way a small studio does.
Grounded generation
Images, video, and audio that stay faithful to the plan and to the facts — generated as evidence of a stated intent, not as free-form decoration.
Iterative editing
Critique-and-revision loops where the agent watches its own output, compares it against the brief, and repairs what misses — the way an editor works with a director.
Provenance & disclosure
Every artifact carries its sources, its edit history, and an honest label. Synthetic media should never be mistaken for a recording of reality.
Open problems
Questions we are working on.
This is a research program, not a product roadmap. These are the problems we expect to publish on as the work matures.
- 01
How should an agent evaluate whether a piece of media actually communicates its intent, beyond aesthetic quality?
- 02
Can long-form generation stay coherent over minutes of video and thousands of words without a persistent world model?
- 03
What is the right intermediate representation between a plan and rendered media — one that supports precise, local edits?
- 04
How do we make provenance a property of the generation process itself, rather than a watermark applied afterward?
- 05
What evaluation methodology measures communication fidelity — whether audiences understand what the author meant?
How we will do it
Built with limits, on purpose.
Labeled by construction
Media made by the agent is disclosed as synthetic at the point of creation, not retrospectively.
Sources travel with the work
Claims in generated media link back to the evidence that supported them.
No deceptive likeness
We will not build toward impersonating real people, and we publish the boundaries we hold.
Open evaluation
Benchmarks for communication fidelity and provenance will be released with the program's first papers.
- Late 2026
Agent architecture: planning representations and edit loops for multi-step media production.
- 2027
First working previews and an accompanying technical report on communication fidelity.
- 2027+
Open evaluation suite for provenance and disclosure in synthetic media.
Media should be made honestly.