Agreed that domain experts should be making these kinds of decisions, and the right way for an agent mistake to get caught is going to depend a lot on the cost and probability of the mistake. My only point here is that the agent side of the system needs to learn. The software production machine needs to improve over time.