Assign the series proactive AI agent teams

Go back to the cliff-walk case. A parent asked whether their children, six and eight, could manage the walk and how to get back to the car. Your answer depends on one fact: the last shuttle leaves the far end at 5:30.

That fact doesn't sit in one place. It's in:

• the TikTok short's on-screen text
• slide four of the Instagram carousel
• a YouTube chapter title
• the pinned comment
• the Facebook group post
• your own voice, halfway through the YouTube walk-through

Now suppose each of your proactive agents does its job well. The research agent checked the timetable on Monday. The caption agent copied the script exactly. Then, two days before posting, the park moves the last shuttle to 4:30. Every agent did its job, and the wrong time still goes out in six places, with your face beside it.

The problem isn't any single agent. It's that nobody owns all the copies. So in this capstone you organize agents into teams that all work from one shared record: every claim, its source and date, and every version that uses it. A shared record is the one place where the whole series agrees on what's true and what has been approved.

How do you decide what the teams are? Evan Spiegel, Snap's CEO, faced a similar problem when a burst of AI experiments needed order. He has described how Snap listed the jobs to be done for Snapchatters (download the app, add close friends, use Lenses) and for advertisers (join the ad platform, set up a campaign). That list showed where agents fit. Snap then built teams around each job and tracked progress against outcomes.

A job to be done is a result someone needs, written without reference to any tool. For your series, the jobs are:

1. Answer the viewer's real question with checked facts.
2. Turn that answer into true versions for each platform.
3. Get each piece released with your approval, and learn from it.

Each job gets a team, a human owner and a measure. Keep the measures about the quality of the work, not about reach. "Every claim has a source and date" is a good measure. So is "every version matches the record." "The short gets 50,000 views" is not, because no team controls that, and chasing it pulls the series toward provocative wording.

Your series runs on three teams. Each one runs its agents in OpenAI Dots or Claude Duet. (You'll choose between them later in the course.) Each team has one main deliverable.

Research and Editorial owns the question, the sources and the claims. Its agents might include an intake agent that groups questions from all five platforms and a source agent that builds packets from approved sources. Deliverable: a claim packet, meaning each claim with its link and date, plus any open questions marked as unknown.

Production owns scripts, visuals, captions and exports. Agents might include a script agent that drafts from the packet and your voice guide, and a versions agent that prepares captions, carousel text and chapter titles. Deliverable: scripts and exports for each piece, each one labeled with a version.

Release and Learning owns the release packet, drafted replies and lessons for next time. Agents might include a release agent that assembles exports, sources and disclosures for your review, and a reply agent that drafts answers to comments. Deliverable: a release packet you can approve, return or hold, plus notes on what to change in the next batch.

The agent names are yours to choose. What matters is that each agent belongs to exactly one team.

Steven Bartlett, host of The Diary of a CEO, has said that every hiring conversation at his company now starts by asking why an agent can't do the job. That's a useful question, and this course adds the reverse one: which person does each agent answer to?

A human owner is the named person who receives a team's output, sends back failed work and makes the decisions the agents can't. In a small creator team it might look like this:

• Editor: owns manual edits and exports in CapCut, so in practice owns most of Production's hands-on work.
• Manager: answers brand and sponsor questions.
• Creator: owns the facts, the voice and the release, so is the final owner of Research and Editorial and of Release and Learning.

On a one-person channel, every owner is you. Write each role down anyway: "Production owner: me. Escalations go to me on Tuesday and Thursday mornings." When a check fails at 11 p.m., the agent needs to know whose queue it lands in, and a written role is easier to hand to a part-time editor later.

An authority limit is a written line that says what an agent may not do, however capable it is. Four limits apply to every team in your Factory.

1. No invented experience. No team writes "my nephew loved it" or "we made the 4:30 easily" unless you lived it and said so. Your observations come from you.
2. No final actions. No agent publishes, sends replies, signs, pays or deletes. Agents draft and prepare. You act.
3. No altering your voice or face. No agent synthesizes or edits your voice or likeness. YouTube's disclosure policy actually exempts cloning your own voice for voiceovers. This course is stricter: if a spoken line is wrong, you re-record it or cut it.
4. Minimal personal data. The record keeps only what the answer needs and no details about minors. Store the question ("Can children six and eight manage the 300 steps?"), not the family's names, messages or the children's details. Agents handle the family's question, not the family.

Paste these four limits into every agent's standing instructions, word for word.

Three teams can each meet their own measure and still pull the series in different directions. So before any work starts, write three things that every team follows.

Audience promise: one sentence on what every piece will honestly give the viewer. For the cliff walk: "Each piece answers a real viewer question about planning the walk, with times and routes checked against the park, and my own experience marked as mine."

Evidence standard: the rules a claim must meet to enter the record. For example: approved sources only (the park's route page and timetable), a link and date on every claim, observations marked as the creator's own, and a newer source that contradicts an older one holds every dependent item until you decide.

Decision record: a running log of every choice, with its date and the person who made it. For example: "Out of scope for this batch: café, dogs, winter route. Decided by the creator." Or: "Shuttle time confirmed from the park timetable page, link attached."

The decision record is part of the shared record. When someone asks later why the chapter title says what it says, the answer is in the log.

Now put it together as a one-page operating plan for your series. Use your own three to five related audience questions, or use the cliff-walk case.

Work through it in this order:

1. List the series' jobs to be done.
2. Group them into the three teams.
3. Give each team named agents and one deliverable.
4. Write each team's authority limits, including the four shared limits.
5. Name a human owner and a work-quality measure for each team.
6. Write the audience promise, evidence standard and first decision-record entries.

A filled-in row might read: "Research and Editorial. Agents: intake, source. Deliverable: claim packet. Limits: approved sources only, no invented experience, no minors' details. Owner: me. Measure: every claim has a link and date."

Keep it short enough to paste into each agent's standing instructions. In the next lesson you'll add acceptance checks: the tests each handoff has to pass before the next team takes the work.