Browser Agent

Tried building a browser use agent from scratch in Python. Other browser use libraries kept breaking at simple steps so I wanted to investigate the bottleneck.

When I compared my solution to the best libraries, they surprisingly had near exact same failures, speed, and task completion rate.

Maybe there is a fundamental issue with how LLMs understand websites and maybe another abstraction or modality is needed to solve this problem.

A good future step is to let Claude Code run in the background to experiment with different ways to expose browser state to LLMs and develop a new language.