Field noteBy The Relmz team
Why we built an AI CMS agent (and the constraints that kept it sane)
The CMS agent isn't a chatbot bolted onto a page builder. It's a streaming tool-call loop bounded by a typed manifest. Here's what changed in our heads.
Drag-and-drop builders age fast. Every editor settles into a routine where the same five blocks land on every page. We wanted an authoring loop that started with the goal — 'a hero with a 6-second loop, three feature cards, a pricing block' — and assembled the page from there.
The trick was constraining the agent against a typed block manifest. Each block declares its prop schema; the agent can't hallucinate components that don't exist, and it can't put a hero inside a hero. Safe by construction.
The second trick was making the edit log the tool-call history. Every move is replayable; every checkpoint is a real version. Drag-and-drop still works — but most pages now start as a paragraph of intent, not a blank canvas.