this is essentially my view, and it's already feasible today with the Code(x)-like environments. It's a question of tuning the harness so that it is more inquisitive (makes fewer choices autonomously). Errors and omissions will be inevitable in expectation (but perhaps they always were)
