The operator changed. The software hasn't.
Every creative tool on the market assumes a human is driving — a human clicking through panels, holding the state of the production in their head. We think the next studio is operated by a coding agent, and that changes what the platform has to be.
The operator inversion
For twenty years, “product” meant a dashboard: screens a human clicks, wizards a human completes, state a human remembers. Coding agents invert that. An agent doesn't want screens — it wants context it can read and actions it can take: an MCP server for state, a CLI for verbs, structured output it can parse, and docs written for machines.
The inversion doesn't remove the human. It moves them to the two places they actually matter: defining intent and reviewing the result. Everything between — the boarding, the queueing, the retrying, the bookkeeping — is operator work, and the operator is now an agent.
A pile of tools is not a studio
You can already hand an agent a box of endpoints: a render MCP here, a filesystem MCP there, a database for notes. It will make you a clip. It will not make you a show — because every endpoint is stateless, so the agent re-describes the protagonist in every prompt, and the lead is a stranger by clip three.
Serialized work — a season, a channel, a universe — is a state problem. Who is this character, exactly? Which take is canonical? What did shot S19-004 depend on? A studio is precisely the thing that remembers.
Identity is a primitive, not a prompt trick
The hardest state to keep is visual identity. Our answer is the canon registry: characters, scenes and props are first-class entities with refs like @mara.v3, locked once and bound to every render that depicts them. A deterministic continuity Advisor scores each render against the locked reference and gates the cut — drift gets flagged, re-bound and re-rendered, not shipped.
That is the difference between “the model was in a good mood today” and the same lead carrying 214 shots at 98% consistency.
Legible to machines, reviewable by humans
An agent-native platform owes its operator a manual it can read — ours is served as fetch-docs and /llms.txt — and owes the human an audit trail they can trust: a machine-readable activity log of every action the agent took, plus a review queue for the calls only taste can make.
Even pricing is part of the contract: metered on what actually costs (render minutes, storage, model gateway) and published machine-readably, so an agent can budget a season before it spends a cent.
Open source, because operators need trust
You don't hand your studio to a black box. The whole platform is Apache-2.0 — self-host it with one docker compose up, bring your own model keys, read every line of the pipeline your agent runs. Hosted cloud exists for the teams who'd rather not run infrastructure, not as the only way in.
Where this goes
OpenMontage lets an agent make a video. lumilib lets an agent run a show — and a show is just the first shape. Any serialized creative work with persistent entities fits the same seven primitives: story, screenplay, canon, storyboard, render, edit, publish.
Bring a story. Your agent runs the studio. You keep the taste.