Assign the campaign proactive AI agent teams

You already run single agents for recurring jobs, such as intake, copy, template and checking. Each one works on its own. So why can a launch still go wrong?

Allie K. Miller, chief executive of Open Machine and formerly a machine-learning lead at Amazon Web Services, separates two roles. A delegate does work for one person and hands it back. A teammate lifts the whole system. She says the best users treat AI as a "first class teammate".

Your single agents are delegates. Each one hands its work back to you, so you become the only link between them. Take the bakery template launch. The copy agent knows the bundle has thirty templates. The template agent holds the layouts, and the checking agent sees the exports. Nobody owns the connections between them. When a fact changes the night before launch, it reaches some files and misses others.

Melanie Perkins, Canva's chief executive, says separate, guarded lanes for product, design and engineering were never a good way to build. A campaign works the same way. In this capstone you'll build a Factory: supervised agent teams that all work from one shared campaign record instead of three separate silos.

The Factory has three teams. Each one regroups agents you already run and produces exactly one deliverable.

Brief and Copy owns the facts and the wording. Your intake and copy agents sit here. It delivers checked copy: every line tagged with the fact it came from.

Design and Video owns the layouts, listing images, preview video and captions. Your template agent sits here, along with any agent that prepares CapCut packets. It delivers versioned design requests and exports, and keeps the editable Canva master or CapCut project beside each one.

Release Review owns the format manifest and the look of the whole set. Your checking agent sits here. It delivers an item-by-item list awaiting approval.

In the bakery launch, Brief and Copy settles that the bundle holds thirty templates and writes the pin headline. Design and Video turns that copy into a pin, post, story, what's-inside grid and preview video. Release Review checks that all of them match and lists each item for a person to approve.

One deliverable per team keeps the handoffs clear. You always know what the next team should receive.

Every team has one named human owner, and you write that name into the record. The owner answers questions the agents can't settle, accepts or refuses work coming in, and is accountable for the team's deliverable.

Marc Benioff, Salesforce's chief executive, says his company didn't build its platform on the belief that AI can do everything, because AI isn't accurate enough yet. Real customer success, he says, takes "humans and agents working together." Use that as your division of labour: agents carry production, and people carry accuracy and release.

In a shop of one, you own all three teams. Write each role down separately anyway:

• Brief and Copy owner: you
• Design and Video owner: you
• Release Review owner: you

Separate rows matter because each role asks a different question. As copy owner, you ask whether the facts are right. As design owner, you ask whether the file reads at phone size. As release owner, you ask whether the whole set agrees. If you hire a part-time assistant later, you can hand over one row without rebuilding anything.

The shared campaign record is one spreadsheet that every team reads from and writes to. It has five parts.

1. Facts, with a version number: the count, sizes, what a buyer can change, claims and dates. For example, Facts v1: 30 templates, square and story sizes.
2. Copy versions, each line tagged with the brief and the source fact it came from. A source tag is that label. For example, Pin headline v2, from Brief A, fact: count.
3. Editable links to the Canva masters and template links, plus the CapCut project.
4. Exports, each recording its own file version and the record version it was made from. For example, Pin_v2.png, made from Facts v1.
5. Approvals: who approved what, which version, and when.

Alongside it sits the format manifest, a list of every item in the sample: pin, post, story, what's-inside grid, preview video and captions.

Why record the facts version on each export? Because later you'll need to find every file made from an outdated fact. A record that only says "pin: done" can't tell you that.

In the Intensive, you gave every agent action one of three settings: always allowed, needs approval or never (never automatic). The Factory keeps them, and every team uses the same settings.

• Filling a draft is always allowed. Brief and Copy can draft pin copy, and Design and Video can fill a layout, without asking.
• An export needs approval. Before Design and Video exports the story at final size, a person says yes.
• Publishing and deleting a master are never automatic. No agent on any team posts the pin or removes a Canva master.

One word needs special care: ready. When Release Review marks the post "ready", it means the evidence exists. The file is opened, checked against the current facts and listed with its versions. It never means an agent may publish. The item goes onto the list awaiting a person's approval.

Write these settings once, at the top of the campaign record, so every team reads the same rules. If a team needs an exception, the owner writes it there too.

The course teaches two proactive agents as equals: Dots from OpenAI and Claude Duet from Anthropic. Pick the one you already have access to. Both work the same way at the level that matters here. Each runs on its own cloud computer, connects to your apps and keeps working toward a goal. It also checks with you before sending, paying, signing or deleting.

Your job is to make it follow the record across team handoffs, so it doesn't improvise.

• Tell it to read only the approved record version, for example Facts v1 as approved, and never a draft row or its own memory of the facts.
• Tell it to return three things: checked copy, versioned design requests and items awaiting publication approval.
• Give it the action settings from the record.

For the bakery launch, a goal might read: "Using approved Facts v1 only, prepare checked copy, design requests and a list of items awaiting approval for the six items in the manifest. Ask before any export. Never publish."

How each product handles app connections, schedules and approval prompts may change. Check these in the current version before you rely on them.

Before any team starts, check two things: which data goes where, and what you're allowed to use.

Connect only the apps you've approved. Your chosen tool runs on a cloud computer, so anything you connect becomes reachable. Keep customer lists, private campaign data and identifying images out of any cloud tool you haven't approved. For the bakery launch, the record needs the template count. It doesn't need your buyer email list, so leave that list out.

Update the permissions list. This is the list you started in the Intro. It records what you may use and on what terms, for mockup photos, fonts, stock and generated elements. A generated element is not automatically cleared. A licensed mockup photographer is credited and paid. If Design and Video is about to place a new stock photo in the what's-inside grid, it goes on the list before the grid is built.

These checks come first because agent teams move fast. If an unlicensed photo or a customer list gets into the run, it can end up in several files before anyone looks.