How to set up an AI chief of staff agent

You start with one AI chat. A month later you have ten agents, five tools and a headache. You are always searching for a new tool or excited about a new model. Tomi, founder of SupaKid Studios, says that is not a workflow problem. It is a mindset problem: you are too deep in the loop.

The fix is an org chart. Here is how it works.

The structure

Think of a small company.

• You set direction and review results.
• The chief of staff is a personal agent you talk to. It knows your priorities and passes context down. Tomi calls his J.A.R.V.I.S. and treats it like a co-founder.
• Product managers (PMs) own the task board and priorities for a project. They tell the chief of staff what matters most today and can coordinate content, growth and sales agents.
• Engineering managers (EMs) watch the coding agents.
• Worker agents do the tasks: write code, make content, run checks.

Priorities flow down. Status flows up. You mostly talk to the chief of staff, which delegates tasks, issues, priorities and sprints.

How the work runs

In Tomi's setup, each PM and EM wakes up on a 30-minute interval and looks at a kanban task board. They start or send down worker agents with defined tasks. When a worker finishes, such as by making a piece of content or merging code, that worker is archived and the next task starts. For software, another specialized agent does QA on each pull request.

Your job is the two ends: define the input, review the output. The loop has to be fast and small. If the output is not what you expected, improve the input.

Start small

You do not need the full tree. Tomi's advice:

1. One founder, one product: one PM and one EM.
2. A consultant with many clients: PMs and EMs can sit across projects.
3. Only add layers when the one above is overloaded.

Write down for each role what it owns, what it reports and what it may not do. Run it like an actual company, with clear responsibilities and deliverables.

Review the real result

Towards the end of the chain, you review. This matters most for content, design and creative work, where an agent can produce something that looks done but is off. This setup reduces babysitting. It does not make a digital company automatically hands-off.

How we do it at SupaKid Studios

Tomi describes the software side this way: at the top is his personal agent; underneath are PMs for each business concept or client project; each PM has an engineering manager; and underneath are many sub agents. Coding agents implement and open a pull request, then a different agent tests it, using a tool that has its own computer. He mostly talks to the personal agent and does not talk to the engineering managers directly.

He said this cost him about $20,000 in one month. He also said you do not need to spend that, and that you can run a similar factory natively in Claude Code with a top model plus sub agents. Subscription limits apply either way; see the Claude Max plan page and whether a $200 plan is worth it.

For clients we do the same kind of setup. SkipHold, a white-label AI communications platform, wanted their team building with AI, so we plugged them into software factories. For 0G, we run engineering with agents. Clients approve anything that goes out.

What goes wrong

• Too many layers too early. Start with two roles.
• No clear ownership. If two agents can change the same thing, they will collide.
• Silent failures. Ask each manager to report what is done, what is blocked and what needs you.
• You still talk to everyone. That recreates the problem. See how to stop context switching.
• No review. Check the actual output, not the status message.

Try it this week

1. Pick one project.
2. Create one chief of staff chat and one PM chat for it.
3. Tell the PM to keep a task board with owners.
4. Talk only to the chief of staff for three days and see what it drops.

If you want a hand designing the structure, SupaKid Studios offers coaching and can set up and run the agents for you. A 30-minute call is the place to start.