What I've found is that even the best models can write very good code "in the small", and the code is mostly correct, but they can be highly problematic in their design, architecture and abstractions they use. The most common problem is that there is a strong tendency to overcomplicate things, and often this can be traced (ironically!) to the agents not having enough context.
E.g. a very common failure mode is that agents are unaware of some invariant in the architecture (or the fact that they could potentially stipulate a new invariant) and write code to handle a dozen edge cases that would never occur. The code is correct, but there's a ton of unnecessary slop.
My suspicion (confirmed by friends and posts on HN) is that at most places, people don't bother reviewing the code and blindly ship all this slop. The reason things haven't imploded yet is a) the code is usually technically correct, and b) all that additional complexity is never exposed to humans -- it is just outsourced to even more agents to manage. Which explains Anthropic's exploding revenue, I suppose.
That said, it's not like things were better before AI. I once had to go Indiana Jones-ing into a humongous legacy Python codebase so full of magic that there were some code paths I could never figure out even with a debugger. I could only tell that if I got a row written to a specific table, some completely disconnected code path elsewhere would get triggered. I suspect what was happening was a function name was being constructed out of 3 different strings from 3 different sources, one of which was a column in the DB row that was getting written, and that function was being eval'd in some thread somewhere.
This was to fix a bug, which I managed to do without ever figuring this flow out. I still spent an inordinate amount of time trying to figure this out before I realized I was being too idealistic and closed the ticket and moved on. I bet an agent would have figured it out though.