Define the three proactive AI agent teams

In the Intensive you built single agents, each with clear stops. A Factory is the next layer up. It is three supervised proactive AI agent teams that share one scene record for a short sequence. A proactive agent is one that wakes on an event and acts without being prompted each time. A scene record is the single, versioned file every team reads from and writes to.

You can use your own three-to-four-scene packet or the sample, scenes 41–44. In the sample, scene 42 has Nora on a farmhouse porch at night, telling her brother she's selling the house.

The three teams split the running paperwork:

• Story and Continuity carries the pages and the takes.
• Production Planning carries what must be scheduled or built.
• Post Coordination carries media and notes.

"Supervised" means people stay in charge. Teams prepare, flag and hold work, and named people make the calls. Think of it as a production office that never wraps, where sign-off still belongs to you.

Each team holds several roles, and each role keeps the four edges from the Intensive: what wakes it, what it may change, what it must never touch, and whom it asks.

Here is one way to staff Story and Continuity:

• Script tracker: wakes when a revision lands. Records which draft the script supervisor has confirmed.
• Breakdown keeper: updates only the breakdown entries a confirmed revision affects, and cites that revision.
• Continuity logger: records take details (props, light, eyelines) and lists known problems.

Take the breakdown keeper. It wakes on a revision the script supervisor has confirmed, not on any file that turns up. It may change the breakdown rows for the affected scenes. It must never touch the take selection, the shot list or another team's deliverable. When something is ambiguous, it asks the script supervisor.

Planning and Post get the same treatment. A planning role might draft the schedule; a post role might organize media. The rule doesn't change: a role with no written edges isn't ready to run.

A team with no output just drifts, so each one produces real documents taken from the course outputs.

• Story and Continuity: scene breakdown (XLSX) and continuity plan (PDF).
• Production Planning: shooting schedule (ICS), call-sheet draft (FDX), location constraints and prop requirements (CSV).
• Post Coordination: edit outline (PRPROJ), media organization, sound cue notes and format requirements.

These are coordination drafts. They show whether departments are working from the same facts. They are not a production-ready schedule or a real budget, and the Capstone doesn't promise either. A shooting schedule from Planning, for example, lays out the scene order and the questions the producer needs to answer. It does not commit a day, a crew or a location.

When you set up your packet, write each team's deliverables at the top of its section of the scene record. Name the file and its format. Then anyone opening the record can see what each team owes and whether it has arrived.

An acceptance condition is a test a deliverable has to pass before anyone treats it as done. Write each one so that a person, or another agent, can answer yes or no just by looking.

Five conditions apply to every team:

1. It cites the script version it was built from, for example "blue pages, confirmed by the script supervisor".
2. It lists its accepted inputs, meaning which files and approvals it used.
3. It keeps explicit script facts apart from implied requirements. "Night, porch" is in the script. Wet steps, a porch light and rain sound are implied, and they don't become production requirements until someone accepts them.
4. It flags missing media rather than substituting another file.
5. It protects what the scene can't lose: the unplanned look back at the house in take six. Nothing in the deliverable may drop it, replace it or weaken it.

You can add team-specific tests on top. For example, the edit outline must show selected versions in order, with generated material labeled apart from photographed material.

Some facts look like paperwork but belong to people. A team never produces availability (who or what can be booked, and when), costs, or safe operating instructions (how to do something safely on set). An agent that guesses any of these creates a number or a procedure that looks real and isn't.

These items appear in a deliverable only as questions addressed to their owners.

Say Planning needs to know whether the farmhouse can be used again for a pickup. It doesn't write "available Thursday". It writes: "Q for producer: is the farmhouse available for a pickup on scene 42? Blocks: schedule rows 42A–42C." If a rain effect needs a safety plan, the deliverable says "Q for the department lead responsible: safe operating procedure for the rain effect". It doesn't draft one.

The question goes in the scene record with the work it blocks. That work stays held until the owner answers, and the record logs who answered and when.

For each team, list the decisions it can't make and put a person's name beside each one. Write roles and names in the scene record, for example "Producer: [name]", so nobody has to guess on a Friday night.

• Script supervisor: confirms which draft is current.
• Director: creative calls, such as which take stays and how a scene plays.
• Producer: scheduling, resources and cost.
• Department leads: decisions inside their department, such as art on props and sets, and post on media and deliverables.
• The performer, with the director: any call about the performer's likeness, such as altering or relighting their face.

Agents may prepare, flag and hold work. They never make scheduling, resource, release, creative or likeness decisions. If a team can't name an owner for a decision, the team isn't ready to run. Leave that decision blocked until someone takes it.

The Factory runs on one proactive agent platform. You have two options, and the course treats them as equals.

• Dots (OpenAI) runs on its own cloud computer connected to your apps. It checks with you before it sends, pays, signs or deletes anything.
• Claude Duet (Anthropic) does the same job.

The course doesn't compare them. Choose on two practical questions:

1. Can you access it? Your organization may already use one.
2. Can you connect it to your scene record? The teams have to read from and write to the same versioned record. If the platform can't reach it, the Factory can't work.

Write your choice into the scene record in one or two sentences, for example: "Claude Duet: our office already has access, and it connects to the shared drive that holds the scene record." Then turn on the platform's check-before-acting behavior for anything that sends, pays, signs or deletes. That keeps a person between the teams and anything that leaves the office.

Four people from the industry help frame the Factory. Each has a commercial interest, stated here.

Ryan Kavanaugh, founder of Relativity Media, organizes a whole production around one performance capture on his company's AI gray stage. He's promoting his own system, and his schedule and cost figures are his own and unaudited. The point to take: center the paperwork on one capture.

Jon Erwin, speaking in his own studio's presentation, describes performing, generating and editing at the same time. When the cut asks and production answers, the paperwork has to keep pace, and each person still keeps their own call.

Bob Iger, speaking beside OpenAI's chief executive, presented Disney's deal, which licenses characters but excludes talent likenesses and voices. So the question isn't whether to use these tools. It's how, and on whose terms.

Natasha Lyonne, co-founder of Asteria, a studio built on a licensed model, argues that artists need a seat at the table. She also calls for attribution and regulation, so don't read her as an uncritical supporter.