LLM systems start to Frankenstein things together at a very modest amount of code. The lack of abstractions cause huge and glaring problems for anything that needs to run predictably.

Abstraction is an issue for every AI coding agent. They are painfully useful, but need constant badgering in the right direction.