Interesting approach. How are you handling the DOM processing inside the sandbox without spiking CPU usage? If it's not making constant API calls, is the vision model running locally (WASM/WebGPU), or are you using a clever way to diff the page state before sending it to the LLM?

Yes, it is kind of a clever way to diff the page state. The ai agent has some built in tools to deal with this. The user gives a prompt say "Watch for X words" The LLM then runs the provided tool with the necessary args. The tool then runs a python loop to check for it in the DOM while the LLM sleeps. Then once it's found the LLM is awoken. Also, there's a tool for watching for changes in pixels in certain regions. It works in a similar way.