Last week I wrote about the PM. The argument: the pile of specs, decks and PRDs that fills a PM's week was never the job — it was the residue of it. AI is dropping the cost of that artifact toward zero, and what's left standing is the understanding underneath. This is the same move, one role over. The question haunting architecture teams has the same shape: if the model writes the service, draws the diagram, and drafts the ADR — what's left for the architect? It's the wrong question, and the confusion behind it is the same one. "Architecture" always meant two things at once. Structure as object: the diagram, the well-factored code, the ADR — the thing you point at and hand off. And architecture as practice: deciding what matters, reading trade-offs, choosing when to decide. We treated them as one for years because producing the object was expensive. And because it was expensive, we mistook the cost for the value. The artifact was never the value. It was a byproduct of the practice — the same residue the spec pile was for the PM. We just never got the byproduct for free. Now we do. And that's all AI changes. @martinfowler and Erik Doernenburg wrote in 2003 that most of the architect's traditional work should be done by tools, or not at all. It sounded like provocation then. Now it's just a description of the tooling. Strip the artifact away and what's left is exactly what the canon has pointed at for twenty years, and no model does: the trade-offs @neal4d and Mark Richards built two books on, @ghohpe's idea that architecture is selling options under uncertainty, @al94781's continuous stream of decisions. None of it is generation. It's judgment. What agents actually change is where intent has to live. It used to sit in people's heads, enforced by code review and the tacit norm that the UI layer doesn't reach into the database. An agent doesn't absorb that over six months of PRs. It does what the context tells it, fast and at volume. So intent has to leave people's heads and become executable — and that's where fitness functions stop being a nice-to-have and become the primary control surface. The diagram became cheap; the fitness function became load-bearing. And the counterintuitive part, also from @ghohpe: an option is worth more when volatility is higher. AI is the highest-volatility environment software has seen in a generation. Which means keeping decisions reversible is worth more now, not less. So the artifact-architect is being automated out. The practice-architect is worth more than they were two years ago. Same title, opposite trajectories. AI cut the cost of saying things, not the cost of knowing them. The mind that knows which diagram to draw, and why, and what to keep changeable for when it turns out wrong, isn't being automated. It's being promoted. Full argument: medium.com/@dsbraz/promoted-…