Two years into working with AI agents daily, my workflow looks nothing like autocomplete. It looks like delegation: I own taste and constraints, the agent owns the long tail of implementation.
The rule that changed everything: never review generated code you couldn't have written. Review outcomes instead — run it, poke it, break it. Taste is the last non-automatable skill.
// the spec is the code now "hero: one 3D scene, lazy-loaded, static fallback under reduced-motion, zero CLS"
The developers who struggle with AI tools are usually fighting for control of the keyboard. The ones who thrive fight for control of the spec.
You are not typing less. You are deciding more.