Governance in an AI-native world: an agent that decides before it builds, a front door anyone in the company can use, and a person who still makes the call.
When a component costs minutes to make, the risk isn't a slow design system. It's a sprawling one. On 28 August I asked who approves a component, and how you stop AI inventing patterns instead of reusing them. This is a first answer: the job moves from making to deciding.
The agent. Before it builds anything, it searches the system and makes one call: reuse, extend, build new, or decline. It writes the reasoning down the way a design system lead would, drafts anything it builds with stories, docs and accessibility notes, and never merges its own work. Three real requests got three different answers: a status label (reuse: the Badge already does it), a checkout error summary (extend: the existing one looked right but its links went nowhere), and a confetti countdown for one sale (decline: campaign work, not system work). Generating code is easy now; saying no well is the job.
An interface for anyone. The agent runs in Claude Code, but asking for a component shouldn't need it. So there's a request desk anyone can use: ask in plain words, get a plain answer back. The design system team sees the same request from the other side and approves, declines or asks a question.
Three rules so far: check for reuse before generating, write the reasoning down, and keep the decision with a person. AI can draft the component. Deciding whether it belongs is still governance, and still a person's job.
Generating code is easy now; saying no well is the job.