Key Takeaways

  • WebCodecs and GPU acceleration now let browsers process professional video codecs in real time, turning the browser tab into an engine as capable as a local FFmpeg build.
  • Kino founder Luke Igel chooses LLMs controlling standard tools like Blender, RenderMan, and video timelines over black-box text-to-video diffusion models every single time.
  • Generative models fail creative directors because they lack precision; coding agents manipulating software primitives give creators deterministic control and clean project files.
  • Igel and his colleague Donald tested this model by sharing a Figma-style mega timeline with autonomous agent swarms running tasks simultaneously.
  • The prompting layer is disappearing into raw context pipelines, with Igel feeding Granola meeting notes directly to the editor.

WebCodecs Turned the Browser into a Real Video Engine

For two decades, professional video editing stayed trapped in native desktop wrappers. If you wanted frame-accurate scrubbing, multi-track audio, or raw camera footage decoding, a browser tab choked. WebCodecs changed that math completely.

As Igel explains, “webc is what allowing this formulation of our business. It is like ffmpeg in the browser fully GPU accelerated if you want.” Modern browser standards allow Kino to run custom decoding for pro camera files directly on the client GPU. That removes the lag that killed previous web editors. It also makes real-time, multi-user timelines possible without streaming heavy video proxies back and forth across a cloud server.

Coding Agents Beat Black-Box Diffusion Models

Silicon Valley is pouring billions into generative text-to-video models. Igel believes that money misses what working artists actually need: control.

When a director asks for an edit, they want to move a camera three inches left, lengthen a cut by twelve frames, or adjust a color grade. A diffusion model cannot do that without regenerating the entire frame, shifting every unintended detail in the process. An LLM acting as a software engineer can simply write the script that changes the parameter.

“If I could choose between some magical lab who could give me a model, a diffusion model that could make a Pixar-like movie or a coding agent that could use RenderMan and Blender and video editors to make a Pixar-like movie, I would choose the LM every single time,” Igel says. Models like Opus 5.5 can already output code for complex motion graphics, assemble live-action footage, and control layout engines. By driving existing production software instead of guessing pixel values, the agent leaves an editable trail of assets behind.

The Timeline Becomes a Multi-Agent Canvas

Once the editor lives on the web and code controls the cuts, editing looks less like Final Cut Pro and more like multiplayer Figma. Human editors no longer work sequentially in isolation.

Igel describes testing this setup with his team: “Donald, my colleague, and I, we spent the past few days living in one giant editor timeline, Figma style, while swarms of my agents and swarms of his agents all just chipped away at the same mega timeline.” One human sets the high-level pace, another reviews incoming sequences, and multiple background agents generate motion graphics, trim silences, and test variants at the same time.

Even the prompt box is fading out. “I don't even prompt anymore,” Igel notes. “I just feed my Granola link to Kino.” The context from the meeting doc sets the project parameters, and the agent translates those notes into timeline cuts.

What to Do With This

Stop building wrapper apps around raw generative video prompts. If you are developing AI tools for creative work this week, map the software tools your users already touch, inspect their CLI or scripting APIs, and build an agent that emits those commands instead of raw pixels.