yeah no I agree, it will need a lot of beforehand set up to work correctly, something with future support of multimodal inputs would be much easier.

yeah no I agree, it will need a lot of beforehand set up to work correctly, something with future support of multimodal inputs would be much easier. For reference this is the demo that peaked my interest: https://github.com/browser-use/jev-ultrafast. A harness is deployed with a discrete action space and a discrete set of observations in a loop, with a final goal of finding a flight with certain params.