Agents for creative tools.

What changed, and what did not

For most of the history of software, a creative tool was something a person operated. The person decided; the tool executed. Agents change the first half of that: an assistant can now read a brief, plan a sequence, and ask a tool to do things, in a loop, without a person clicking each one. What has not changed is the second half. The tool still executes, it still costs what it costs, and it is still the thing that knows whether a request makes sense. So the interesting question about agents for creative tools is not what an agent can do. It is what a tool should let one do, and how it should say no.

The Model Context Protocol is the standard that makes the question practical. A tool that speaks it publishes a set of named operations — list the productions, read this board, render that frame — with what each one takes and returns, and any agent that speaks the protocol can hold those operations as tools. One connector, every assistant, rather than an integration per vendor. The connect page is the door; this page is about what should be on the other side of it.

The shape of a good connector

Reading is free and should be rich. The most useful thing an agent does with a production is read it like a director rather than like a database. "Every warned frame in scene four with Mara in it" is one question, and a connector that can only return rows by id makes the agent reconstruct the answer from forty calls. PREVIZ's server answers that question as one call, exact and free, because the board, the locks and the drift verdicts are the data, not a layer over it.

Spending asks first. A render costs credits. An agent in a loop can ask for a render a hundred times before anyone notices, and a connector that runs on request is a connector that will one day spend a budget on a bug. Every generating tool in PREVIZ answers with a quote — what it will cost, what it will refuse and why — and runs nothing until the agent confirms. A dry-run quote for a whole plan is a tool of its own, so an agent can price a sequence before it commits to a frame.

The same gates as the room. An agent is not a side door. The project's engine policy, the continuity locks, the approval rules and the provenance record apply to a frame an agent asked for exactly as they apply to one a person asked for. A render an agent made is scored against the lock like any other, placed on the board as a real card, and carries its lineage.

A key carries a budget. Each key can hold its own monthly credit cap, enforced on the server. Handing an agent a key is handing it a budget, and the budget is a property of the key rather than a promise in a prompt.

Two ways in, both honest about who is asking. An assistant that takes only an address signs in and is shown a consent page that names the app, the account and what an approved app can do, spending included. A client that can send a header uses a key. In both cases the work runs as the person, under their role, never above it.

What an agent can hold in PREVIZ

List productions and read a board with its locked identities. Ask for a lane, a cast member, a verdict. Pull the provenance of any asset — what it was made from, by which engine, under which locks. Generate frames against a reference and get back the one that matched and how close the others were. Animate takes from an approved frame. Do surgery on a frame — depth, layer separation, inpainting, relighting — through the same lanes the room uses. Lock an identity into the continuity banks. Browse the look vocabulary. Place any result on the board as a card a person can pick up. And get a quote for all of it first.

Inside the product, the same idea runs the other way. Infinitum is a board with a director agent on it: give it a paragraph and it plans a sequence, holds the locked cast by name, and hands each step to the room that owns the decision, quoting before it spends. It is bound by a written constitution — act, never bluff, never repeat, continuity by default — and its verbs are the same ones the connector exposes. An external agent and the internal one are two hands on one set of controls.

Beside the connector sit the doors for people who work inside other applications: panels for the two edit-bay applications most cutting rooms use, and an add-on for the 3D application most previs departments use, so a board can be reached from a timeline or a scene without leaving it.

What goes wrong

An agent that cannot see the budget. Loops are cheap to write and expensive to run. If the connector does not quote, the agent cannot plan a spend, and the first sign of trouble is the invoice. Insist on a quote tool and a cap per key.

A connector that bypasses the room's rules. A frame made outside the locks is a frame that drifted with permission. The gates must be the product's gates, applied at the server, not a convention the agent is asked to follow.

Results that land nowhere. An agent renders ten frames and returns ten URLs, and a person has to carry them into the board by hand. A result should land where a person works, as a real card with its lineage, or the agent has made a folder rather than progress.

Mistaking access for authority. An agent holds the key of the person who gave it; it does not acquire a role of its own. A connector that lets an agent approve what the person could not is a connector with a hole in it.

The workflow overview places agents among the six stages; the trust page says what is recorded about every generation, which is what an agent's work is also held to. Every plan has the connector, and plans differ by credits alone.

Questions people ask