Assign strategy, resource and execution proactive AI agent teams

A Factory in this course is a set of coordinated proactive AI agent teams that prepare one priority for your judgment. A proactive agent is an AI assistant that runs on a schedule or a trigger and does assigned work without being asked each time. You have built single proactive agents already. Now you organize them into teams. They need one clear initiative to work on.

A bounded initiative has a fixed scope, horizon, constraints and decision date. "Grow into new markets" is an ambition. Teams cannot prepare evidence for it because they cannot tell what is relevant.

Example: a facilities services company of 120 people wants to grow. The bounded version reads:

• Initiative: launch office cleaning in one neighbouring city
• Horizon: 12 months from approval
• Constraints: a budget ceiling you set, no more than one new site manager, current clients' service levels protected
• Decision owner: you, the CEO
• Decision date: go, delay or stop at the quarterly review

Write your own initiative in this format. Use the decision brief from Day 1 if it fits. Keep it to one service and one location so three teams can work on the same thing.

Split the work so each team prepares one kind of evidence. That way no single agent produces an attractive story that hides costs or conditions.

Strategy team. Is the opportunity real, and how should we position it? It prepares market evidence, a competitor landscape, customer themes, segment profiles and positioning options.

Resource team. What would it take, and what can we afford? It prepares the demand model, scenario tradeoffs, the capacity envelope and a role-level staffing plan. The capacity envelope is the range of work your people, budget and equipment can handle before something breaks. Role-level means it lists roles and numbers, such as "2 shift supervisors". It never names individuals.

Execution team. Could we deliver it, and what could go wrong? It prepares delivery milestones, the dependency register, risk controls and the escalation route.

In the cleaning example, the strategy team might find that mid-sized offices complain about missed cleans. The resource team checks whether a second shift fits the capacity envelope. The execution team flags that the new site depends on a van lease being signed.

None of these outputs is a decision. They are inputs to yours.

Give each team a short written charter. It should have four parts:

1. Scope: what the team works on, and what it leaves to other teams.
2. Approved sources: the only places it may read. Name them, as you did for the signal agent on Day 3.
3. Expected outputs: the files it must produce, with their format.
4. No decision rights: it may not decide or commit to spending, staffing, contracts or strategy.

Example charter for the resource team:

• Scope: demand, cost and capacity for the city launch only. Leave market sizing to the strategy team and delivery dates to the execution team.
• Approved sources: the approved budget sheet, current shift rosters with names removed, the supplier price list, and the scenario sheet from Day 5.
• Outputs: demand model (XLSX), scenario tradeoffs (XLSX), capacity envelope (XLSX), staffing plan (CSV, roles only).
• Decision rights: none. Options go to the CEO.

Leave employee details, board papers and acquisition plans out of the sources unless your organization has approved the tool for them.

A handoff is when one team's output becomes another team's input. An example is the strategy team's segment profiles feeding the resource team's demand model. If that handoff exists only inside one team's chat, the other team is working from something it cannot see or check.

The shared record fixes this. Build it on the leadership decision log from Day 6, which already holds approved decisions, owners and review dates. Add a handoff section. Each entry records:

• from which team, to which team
• the output and its version or date
• the sources it relied on
• open questions or assumptions
• status: ready, blocked or awaiting the executive

The rule: a team may only use another team's work if it appears in the shared record. No team relies on another team's private conversation.

In the cleaning example, the resource team picks up "Segment profiles v2, 14 March, sources: client survey and city business register, open question: share of offices needing weekend cleans". It does not pick up a remark the strategy agent made in passing. Next lesson you will set the criteria a handoff must meet before it is marked ready.

The course works with either OpenAI Dots or Claude Duet. Your job is to have one tool follow the approved shared record across team handoffs. The tool should return changed evidence, resource options and decisions awaiting you. This lesson does not compare the two tools' features. Features change, so check them yourself in the tool you choose.

Decide in this order:

1. Access: which tool you can actually use now.
2. Approval: which tool your organization approves for this kind of material. Ask your IT or data owner if you are unsure. If neither is approved for sensitive material, use the sample initiative and anonymized data.
3. Check in the tool: confirm in its current documentation or settings how it reads a shared document or log, how it runs on a schedule or trigger, and how it resumes after an interrupted run.
4. Record the choice: write one or two sentences in the shared record explaining why.

Example: "We use Claude Duet because it is the tool our company has approved for internal documents. Before Day 9 I will confirm how it reads the decision log and resumes a stopped run."

Before the teams start, estimate how much the shared record will hold. This tells you whether the record is something you can actually read at review time.

In the cleaning example there are 3 teams. Their listed outputs total 13: strategy 5, resource 4, execution 4. That is roughly 4 per team. If each output is handed off once per cycle and you run 3 compressed cycles, as you did on Day 7, the record gains about 3 × 4 × 3 = 36 handoff entries.

That volume is manageable only if every entry follows the same short format and shows a clear status. Then you can filter for "awaiting the executive" and "blocked" instead of reading everything.

Move the sliders below. Notice how quickly entries grow when you add a team or extra outputs. A fourth team, or a team producing every file the syllabus lists, adds work that may not add judgment. If the number looks unreadable, narrow a team's scope before you add tools or agents.

To finish this lesson, write in your shared record: your bounded initiative, three team charters, the handoff entry format and your tool choice.