Make a performant WebBrowserAgent class

[aymeric-roucher]

Web browsing is a highly specific task, requiring very different tools and state than other agentic tasks.
There are two big avenues for developing web browsers:

• text-based
• vision-based

At the moment, from internal tests, text works better than a raw vision from a base VLM that has no labelling done to help him click on screenshots.
But browser-use has lots of success with such a scaffolding for vision models, so it could be something to try out.