Agentic Video Style Experiment
Experiments·May 26, 2026
A small experiment from some agent video work, testing whether an agent can research a visual style, iterate on frames, and carry a chosen design into a video model.
Disclaimer: This is wholly derivative of Vox video styles and was made only as a generative video experiment. I would not intend to use this style in my own work. The reference was used here for expediency while testing the workflow.
Like many people, especially in tech, I am still doing a lot of work the traditional way while also experimenting with more agentic workflows. I am going to try to post small experiments as I go. Here is the latest one.
- Agent: researched the styles of Vox videos and defined a reusable style file.
- Agent: sent tests to an image model, reviewed them, then improved the output over two or three rounds.
- User: picked the design.
- Agent: sent the reference image as input to a video model.
The main thing I am interested in here:
- Controllable and reusable style that can get better over time as you make more images.
- An iteration loop where an agent looks at style frames until it is happy.
The next step is getting the agent to review the video output itself. I have done this before with short PNG sequences, and I will probably use a similar approach here so the agent can reason across the actual motion, not just the still frames. Hopefully that catches things like arrows animating the wrong way in the video above.