How a code agent can create AI videos by reusing assets

What if creating a professional video did not mean manually jumping between ten different tools, but asking an agent to coordinate the whole process?

That is what I have been testing in this experiment: a workflow where a code agent does not only help edit a video, but takes part in almost the entire production process. It starts from an idea, turns it into a script, splits the content into scenes, identifies the assets that are needed, reuses assets from previous work, generates the missing pieces and assembles the final result.

The interesting part is not one specific tool. The interesting part is the way of working.

The problem: every video starts too much from scratch

Producing video consistently usually comes with a lot of friction. Even when the format is similar to previous videos, we often repeat the same steps: define the structure, write the script, find references, record screens, generate images, prepare music, synthesize voice, assemble scenes, adjust transitions, render and review.

The problem is not just time. There is also the mental overhead of coordinating very different pieces. A technical video may need a browser recording, an interface animation, well-synchronized narration, music that does not compete with the voice, several supporting images and an edit with rhythm. Each of those parts usually lives in a different tool.

That is why I am interested in using an agent as a coordination layer.

The shift: the agent as orchestrator

In this workflow, the agent does not try to replace every tool. It does something more useful: it connects them.

The agent understands the goal of the video, decides what each scene needs and chooses which tool to use for each asset. If the video needs to show an application, it can prepare an automated recording with Playwright. If it needs music, it can create a track with Lyria. If it needs voice, it can synthesize narration with Gemini TTS. If it needs visuals, it can generate images or clips. And once the assets are ready, it can integrate them into a HyperFrames composition.

The key is that the agent is not asking a "generic AI" to do everything. It decides what kind of asset is needed and uses the right tool for that specific job.

The workflow we tested

The experiment starts with a simple idea: explaining how a code agent can create videos using AI.

From there, the agent breaks the work into phases. First it defines the script and estimates the approximate duration. Then it splits the content into scenes and decides what should appear on each screen. Some scenes need animated interfaces, others need browser captures, others work better with generated images and others only need simpler motion graphics.

Once the storyboard is defined, the agent checks which resources already exist. This part is especially important. Before generating anything new, it looks for intros, logos, transitions, music, clips, screenshots or animations that can be reused.

Only then does it generate what is missing.

Reuse before generating

For me, this is the most powerful idea in the workflow.

Every video you produce leaves behind a small library of resources. It might be a branded intro, a transition, a CTA animation, music, a product clip, a screen recording, a base image or a scene structure that worked well.

If those resources are organized, the next video does not start from zero. It starts from a library that already knows how the brand looks, what visual tone works, which components can be repeated and which pieces should be adapted.

This changes the economics of production. The first video costs more because many pieces need to be created. The second one can already build on part of that work. And as you produce more content, the library grows. Each publication leaves resources that make the next one easier.

It is not only about saving time. It also helps maintain visual and narrative consistency.

Specialized tools for each type of asset

In the video, we show a pipeline where each asset comes from a specialized tool.

HyperFrames handles the composition, timeline, layers and render. It is where everything is synchronized: scenes, narration, music, videos, images, overlays and transitions.

Playwright is used to automate browser recordings. Instead of manually recording an app, the agent can open Chrome, enter a screen, move the cursor, click, open modals, scroll and capture the result.

Gemini TTS generates the male Spanish narration. In this case we ended up splitting the voiceover by scene to avoid a sentence carrying over into the next screen. That detail was key to making the video feel more synchronized.

Lyria generates the music. We created an instrumental track better suited to the rhythm of the video: technological, dynamic, but restrained enough not to compete with the voice.

We also used image models to create supporting visuals for application scenes: marketing, training, enterprise and software. Instead of leaving empty cards or weak mockups, we generated images with an aesthetic that matched the rest of the video.

What we built in this video

The result is an 85-second piece that explains the production workflow itself.

The video starts with a quick hook, moves through the agent's planning, shows a storyboard, presents the reusable asset library, introduces parallel asset generation, shows a browser automation with cursor movement and clicks, includes a real clip of the HyperFrames editor and ends with a cumulative idea: the more content you produce, the more resources you can reuse.

During production, we adjusted several things that are very typical in a real project. We centered visual elements more carefully, gave titles and content more breathing room, replaced empty boxes with generated images, removed a "production mode" block that was not adding enough, recorded the HyperFrames editor, regenerated the music, changed the final CTA and refined the synchronization between voice and screens.

There was also an interesting technical detail: the HyperFrames editor clip needed more frequent keyframes. Otherwise, during rendering, residual frames could appear just before a transition. We re-encoded it so the render would be more stable.

Iterating in natural language

Another important part of the workflow is iteration.

Instead of rebuilding the whole project, the agent can modify only the affected scenes. For example: changing the CTA, adjusting the narration, replacing an image, moving a cut, changing the music or re-recording a specific screen.

That makes revisions much more manageable. The project stops being a closed block and becomes a composition that can be edited in parts.

In this case, the iterations were quite natural: "the narration does not land well on some screens", "the HyperFrames clip shows the previous mock", "the titles are too tight", "change the CTA". Each observation translated into concrete changes in the composition, assets or audio.

Why this matters

I think this approach has a lot of potential for technical content, marketing, online training, product documentation and onboarding.

In all those cases, there is a constant need to produce video, but there are also many patterns that repeat. There are products that need to be shown many times, screens that do not change much, visual styles that should remain consistent and resources that can be adapted.

If an agent can rely on a growing library and coordinate specialized tools, producing video stops being a sequence of manual tasks and starts to look more like a system.

And that is what I find really interesting: not just creating a video faster, but building a process where each video improves the next one.

The question

This video is a test of that workflow.

It is not the end of the road, but it is a clear signal of where agent-based content production can go: less repetitive work, more reuse, more automation and more ability to iterate quickly.

The question now is:

Would you like to know how to create something like this?