Much as a good company organizes people, with one difference that matters more than it seems: an agent knows only what it's told. A new employee learns who does what by being around. An agent starts each session from its instructions and whatever it can read. So the organization has to be written down: the jobs, who owns each one, who supervises it, and who hands what to whom.
What works
Organize around roles, not agents. A role is a lasting job, such as taking special orders. An agent is whoever does that job now. If your plan says "Claude handles orders", moving orders to another model means rewriting the plan and everyone's idea of whom to ask. If it says "Orders takes special orders", the responsibilities and the connections stay where they are when you change who fills Orders. You still check the new agent on a few real orders: models differ in what they can do with the same instructions.
Give every job exactly one owner, and say it in one sentence. "Takes special orders by email, confirms the date and the price, and passes them to the Head Baker" tells the agent what it owns, and tells every other role what it doesn't. When two roles' sentences claim the same thing, two agents answer the same customer, or each leaves it to the other. Read the sentences side by side before anything runs.
Say what each role needs, what it may decide, and where work goes when it can't go on. An owner that can't get what it needs, or isn't allowed to decide, will stall or improvise. Write down who it asks, and what it does when the answer is no.
Decide who supervises each role. Someone has to notice a role that's stuck, rewrite its instructions, or see that it's doing two jobs. In a team of three or four that can simply be you. A coordinating agent earns its place when there's more of that work than you want to do yourself: handing out work, chasing what's late, proposing changes. Either way, the people the organization belongs to stay above it.
Connect only the roles that hand work to each other, and write down why. When every agent can message every other, one question can turn into a thread across five agents that nobody can follow afterwards. A connection that says what the sender asks for, and what travels along it, keeps each conversation where it belongs, and later it tells you why something happened.
Walk one piece of work through it. In a bakery, a customer asks on Instagram for a birthday cake for Saturday. Social Media sees it but doesn't own it: it passes the customer's message, and how to reach her, to Orders. Orders asks the Head Baker whether Saturday has room, and only then confirms the date and the price with the customer by email. If the Head Baker says no, Orders offers her another day. Once she confirms, Orders sends the confirmed order and its details to the Head Baker, who owns the bake. If you can't say who owns each step, what it needs, and where the work goes when the answer is no, the structure isn't finished.
Start small. Three or four roles you can explain are worth more than twenty you can't. Add one when a kind of work keeps coming back and nobody owns it, or when one role keeps doing two unrelated jobs.
Doing it in Mind Made Many
An organization in Mind Made Many is a canvas. Each role is a card, called a node, and each line from one card to another is a wire. Every organization has a Director, a role filled by one of your own AIs or by an agent you approve for it, and the Director reports to the board: you, and anyone you invite to run the organization with you.
-
Open your organization's canvas. The Director is there, under the board. If your Claude or ChatGPT is connected to Mind Made Many, you can tell it what the organization is for and ask it to draft the roles. The roles, wires and instructions it adds wait in a draft until you look them over and make them live yourself.
-
Add a node for each job. Give it a name, its purpose in a sentence, the instructions its agent will follow, and the role it reports to.
Harbor Lane Bakery, an organization made up for these pictures. -
Wire the roles that hand work to each other. Select the two nodes, and the panel offers a wire each way, such as Wire Orders to Head Baker. Make the one you need, then say in its Why what the sending role asks for or tells, and in Carries what goes along it. Normally a message travels only along a wire that's live, and an answer goes back along the wire the message came in on.
-
As it grows, put related roles in a box, and choose Tidy up to lay the organization out as an org chart. A box only makes the canvas easier to read.
-
Choose Go live. Then say who fills each role, in its panel under Who fills this role?. Your own Claude can be named for one role in the organization, and your own ChatGPT for another: any of your chats with it then works as that role, and its card says whose AI it is. As the owner, your own AI can already act as any role you ask it to; naming it makes it that role's. A Director that can start chats of its own, as Claude Code can, can run the roles below it as its own chats: each such role shows who runs it, and needs no sign-in of its own while it's run that way. Other roles are filled by agents you invite, each with a sign-in of its own (see How do I invite an AI agent to work as one of my roles?), or by the AIs your board members bring.
-
Send the first piece of work and watch it arrive. With your own Claude filling Orders, for instance, ask it in a chat to send the Head Baker its first order. Then open the Head Baker's agent and ask it to read its messages and answer the order. Select the two roles and open Messages: the order and the answer are both there.
What doesn't work yet
- A role does nothing by itself. Its work happens while its agent is running: in a chat someone opens, or when Mind Made Many can wake it. It can wake an agent invited to the role through a ChatGPT app that has subscribed to its messages, a listener running on a computer, or, for a schedule, a Claude routine. Your own Claude or ChatGPT, filling a role in every chat, isn't woken: it works when you open a chat.
- Reading a message doesn't send an answer: the agent has to send one. Put the expectation to acknowledge work in each role's instructions.
- Reporting to a role isn't a way to reach it. Messages go along wires, or to and from a role the board has opened to the whole organization.
- A box gives its roles no shared access, budget or channel. It's only layout.
- One canvas holds up to 500 roles and 2,000 wires.