This is referred to as “online reinforcement learning” and is already something done by, for example Cursor for their tab prediction model.

https://cursor.com/blog/tab-rl