I wrote about a different approach to AI agents. Instead of building bots that screenshot your screen and click buttons (like OpenAI Operator or Anthropic Computer Use), I built 304 browser tools where the AI agent IS the app; calling local functions directly instead of navigating UIs.
Everything runs client-side: JavaScript for simple tasks, Rust compiled to WASM for heavy computation, and ONNX/TensorFlow.js models for AI features like background removal.
Full article:
https://kitmul.com/en/blog/building-kitmul-ai-tools-os
Curious what you all think; is the “AI controlling existing apps” approach fundamentally flawed, or am I missing something?