If the human is doing a proper workflow for the AI, I'm not so sure it's big central. Yes, AI needs to be guided strongly, and if it's not, it makes very poor choices/decisions routinely, but that's why you keep an engineer controlling the process.

I force all work to be split up into chunks, properly investigated, spiked, and planned out, and then implemented via TDD (with only system-boundary mocking permitted). It's not perfect, but I've seen regression bugs very much cut so far.

Beyond that, if it's a GUI app, I also require the TDD cycle to require GUI tests with screenshot evidence, and proper assert statements.