I dont want to stretch the analogy too thin but in this case instead of LLMs being a catalyst, perhaps its a monopoly.
My coworker took a class with case studies and the theory presented by that class was that all successful projects have at least one person who has fit the entire system into their head. They can tell you what happens if you pull on this thread. What the consequences are of trying to remove this feature. Lose them and you are fucked. Until or unless someone else steps up and does the same. If they can’t it’s the beginning of the end.
Connect them with clear APIs that don't have to change all that often, and you can build pretty big things.
I imagine this is doable with hardware too
Emphasis on clear. It's a challenging endeavor to properly draw and enforce these service boundaries.
In a system that is the composition of 30-300 different functional units, nobody will be close to any one part unless they’re the bus number for it. So each piece needs to be dead obvious so you can worry about the consequences of composing them. At the end of the day it’s Kernighan’s Law but rephrased so as not to ignore Conway or Brooks.
This is one of the things monorepos help in some ways (by making it easier to change two systems) and break in some ways (as you now get less annoyed by the split between systems being in the wrong place)
In particular it’s difficult to train new people to take up on-call duties if they cannot sit in the corner of the room and try either the same things you’re trying or their own pet theories without taking your attention or interfering with your tests. They need to be able to hear the repro steps and spool up their own snapshot in a similar state. That scales. Gatekeepers do not. Discoverability is necessary but insufficient to achieve this. There’s more to it but the foundation is discoverability and reproduceability.
(It also usually involves some quite proactive learning of the form of finding the specialist in some part of the system and sitting next to them and going "explain to me how this part works", and then repeating more or less indefinitely)
The key there was that one qualified representative for each part of the affected organization and other relevant parties in the core team, and there were frequent alignment meetings where they were all present. The representatives were close to the action, not some three-levels-removed manager.
In the meetings processes and such would get discussed step by step in detail, and any representative could chime in at any point to say "this won't work" or similar.
Hard to do though, and quite costly in terms of organizational resources required.
Or the company ponies up and rewrites the thing.
Sometimes legacy really is bad, and sometimes you really do need to throw the old product into the bin.
See Trello for an example.