One of the problems I see is that the more LLM-driven code becomes, the more voluminous it gets, much of it being completely unnecessary. For example, CC will happily implement new patterns, pull in extra dependencies, and set up event-driven [0] architectures for problems that absolutely do not require any of it.

This does not help the review process. The generated appears to close the ticket, but now you're faced with the dilemma of either LGTM-ing and paying the price down the line or pushing back and arguing with PMs. It doesn't help that sycophantic LLMs hype up any idea and are a breeding ground for confirmation bias.

LLMs amplify. They make the good better and the bad worse. A good engineer will know how to steer the clankers to results an unaided dev could not achieve. And then there is Sturgeon's Law.

0 - this is baffling to me. CC Opus 4.5 and 4.6 will regularly recommend a using custom DOM events to pass information around a trivial React frontend. Same thing on the backend, it's events everywhere, despite no precedents and a clear .md file describing the architecture and philosophy.