My first attempt at this library was to design a minimal API that let me ship shaders with a few lines of code. It was the "dream api" that I always wanted, super minimal, abstracting all the verbosity away, but it came with a cost: It was hard for agents to use It had too many magical abstractions that you have to learn. That is fine for humans (you just learn it once), but my agents kept forgetting them as soon as their context got compacted It also made performance optimizations harder, since I had to "fight the library"

I started going back and forth between abstraction levels to find the right utilities. Keeping low-level control while making it easier for agents to understand. After shipping many shaders on vercel.com, we found that allowed us to improve on performance a lot

Agents usually just open a browser to check their work; waiting ~10 seconds to check for a syntax error is too slow. This is where the CLI helps: - `vpgu docs` and `vgpu examples` to quickly learn - `vgpu check` to validate shader code syntax

If your agent is shipping in a cloud sandbox, it probably won't have a GPU. Usually they go, "oh I cannot render this", and give up. vgpu solves this by shipping a CPU renderer that they can install to test their code. This also makes `agent-browser --webgpu` work anywhere

The headless renderer also lets us write tests that run on CI. ...or render our shaders as high-resolution videos locally. On Mac, rendering runs natively on Metal, so it's real GPU speed

And finally, my favorite feature: WGSL modules. This allows us to write shaders in the WebGPU Shader Language We can now create shader libraries and registries. The WGSL code gets minified and validated at build time.