I think a lot about this and the types of harnesses we need to build in order to support a post-code-review world. It's still all up in the air, it seems.
Definitely.
I think about this has having two separate but related concerns; how do we build tools that focus attention on the parts of the problem where human judgement is highest leverage; and how do we build harnesses that enforce invariants such that humans trust the attention focussing part. It's a challenge!
Some of my personal thoughts on path dependence in the world of zero marginal cost code.