If you can define what correct looks like in the form of failing tests, AI agents can do the work of making them pass. TDD always had the right instinct, but now there's cheap, tireless labor to actually exploit it. The hard part is that writing good tests requires deep domain knowledge, and AI isn't great at that yet. It'll generate hundreds of shallow unit tests that cover nothing meaningful. So the real skill shifts from writing code to writing specs. The implementation is the easy part now.