A studio you assemble
An AI agency in a box.
Not one assistant wearing many hats — a roster you write. Mint the specialists a job needs — an engineer, a project manager, a designer — seat each one on the model that suits it, local or cloud, and hand the room one prompt. They work with real tools until an artifact comes out: pages, images, plans. You approve it, or you send it back.
Not one assistant wearing many hats.
staff a job with two of them or a dozen — each with real tools in hand: the web, the shell, the render pipeline
One room, several minds, a budget.
A job runs in a workspace: the personas you picked, in one room, contributing to the same artifact. Each sits on the model that suits its work — local models on our own hardware for the long grind, a cloud model where it earns its keep; most real jobs use both. And every room has walls: a token budget, a turn limit. A job ends by design, never by runaway.
One prompt. The whole teaser site.
A simple prompt: teaser pages for strata→signal, one hero image each, some mystique — based on earth, computing, strata, and amber themes. Out came the design gallery — the whole teaser site, seventy renders and all — and it is still up. Local models did the work, and every picture still carries, inside its own file, the recipe that made it.
The back rooms.
Behind the roster there is a longer hallway — research benches, lab rigs, test harnesses — the rooms we use when a question needs an experiment instead of an answer. They don’t fit on a teaser page, and they don’t need to.
Parked between jobs, on purpose.
The agency spins up when there is work and parks when there isn’t. Off means off: nothing calls out while it is idle. The gallery is the receipt that the whole loop works; the next job is the reason it wakes.