Counterintuitively, as AI improves some tasks become cheap enough that it's economically viable to employ someone to run them continuously with agents, but until experimentation became low-cost such attempts rarely happened