i was thinking of this too; my ideas were either: 1.

i was thinking of this too; my ideas were either: 1. fully throw jev at a navigation task end-to-end, but at the significant cost of it doing something disastrous or just not getting there correctly 2. let an LLM tool-call jev to navigate sections at a time, but at that point you could also just have the LLM write out batching steps to execute a whole suite of actions in one go (e.g. LLM outputting "move to px(50, 20), left-click for 200ms, move to px(200, 400)"), and you'd be getting the same end result maybe im missing something but i didnt end up figuring out a way to keep jev useful for long-horizon navigation tasks. especially since its intelligence is equivalent to 5.6-Terra non-reasoning, which is quite dumb and is what caused my worries for 1) xd