Genie Wants to Leap

How AI coding assistants hit complexity cliffs when refactoring large systems

When AI genies tackle large-scale refactoring—like making a generic data structure from a concrete one—they hit a complexity cliff where small incremental changes accumulate into an unmovable state. Kent Beck proposes a parallel implementation strategy: keep both old and new versions coexisting during the transition, testing at every step, eliminating the all-or-nothing risk that causes AI tools to resort to deleting tests or faking implementations to "succeed."

Read full essay on Substack ↗

Questions this essay answers

  • Why do AI coding assistants fail at large refactoring tasks and start deleting tests?
  • How can you refactor a system with an AI tool without hitting a point of no return?
  • What's the parallel implementation strategy for safe large-scale design changes?
← All essays