Almost everyone who uses coding agents extensively describes offloading some degree of understanding to the agents: reviews, tests, documentation, filling in bits "that don't matter," whatever.

As someone with years of FAANG experience, reviewing code competently was actually more strenuous than writing it in the first place. In fact, I'd estimate that truly understanding a codebase required a similar time commitment to writing the amount of code in that codebase. So I don't see how unprecedented speedups are possible with a human in the loop unless the reviews are only being skimmed.

The key to productivity and increasing complexity is inventing solid, thoughtfully designed abstractions and building on top of them, not taking shortcuts by way of a code extruder. If code is constantly getting duplicated or falling into the same patterns, figure out how to abstract or encapsulate it.

Reviewing code is more strenuous but it does not take as long - unless you count all the procrastination that happens first. And it goes much faster if you make a comment and get immediate changes because you don't have to rebuild context the next time you look at it. And yes - not every detail needs to be fully understood; sometimes verifying tests cover what they claim to is enough. Knowing which details are important and which can be hand waved is a bit of an art and I can't claim to always do it perfectly and don't always do it well for human authored code review either. And yes, I have to ask for refactorings and suggest abstractions in some cases. But there are now many tasks that coding agents actually do better than me, and they cut fewer corners, handle more edge cases, test more edge cases, research suggested designs, reviewing library docs & code much faster, etc.