9. Ceiling and judgment

Governing a business factory means deciding what must never happen and what is left to each role's judgment. The model in this book answers with an unusual combination: a firm ceiling and almost no mandatory approvals.

The ceiling

The ceiling is the access each conversation carries (chapter 3): the person's, in their owner conversation; the most restricted one, in everything that runs without a person present. And each person's copy contains only the roles they have access to, so that no role outside their access can be reached, not even by mistake. What the access does not allow does not happen. No instruction, skill, memory or message raises it.

For the ceiling to mean anything, each conversation has to prove who it is. No screen and no service can simply say who is acting: the identity service signs a short-lived pass, attached to each call, that states which person or which conversation it is, of which role and with what access. A pass that expires in seconds cannot be reused apart from its call, and even if one part of the system failed, it could not pass itself off as another.

No approval cards

The most common way to control an agent is to stop it before certain actions and show a person a card to approve. The model started out that way, and abandoned it. The reason is that constant approvals added more noise than control, and that a freer way of working was preferable, one in which the roles themselves decide whom to ask for approval.

Instead of cards, each role decides what it needs to ask and whom, and sends it to that person's owner, which is the only one that asks them anything (chapter 4): with the context, a recommendation and, if needed, options. If what it is about to do cannot be undone, the question says so with a mark.

Rules as guidance

That does not mean there are no rules. Whoever administers the business can write approval rules, for the business or for a role («before deleting a customer, ask»; «before changing a price, check with the proprietor»), and some always come built in: before touching credentials, before deleting or doing anything irreversible, before changing what a role is told or allowed.

The difference lies in their nature: they are guidance, not a lock. The role receives them as part of what it knows, as an indication of when to ask. What acts as a lock is the ceiling. The only practical exception is changes to a role's configuration (its instructions, its skills, its rules): those are made by the conversation of a person with access for that, when that person asks, and never by the role's main conversation on its own.

Example. Administration wants to archive an old skill. Its rule says to check with the store manager. It sends the need to the manager's owner, and this appears in the manager's conversation: «Administration is going to archive the skill "Cash close 2025", which is no longer used. The recommendation is yes: the new one replaces it». The manager replies to go ahead, and Administration archives it.

The risk that is accepted

This scheme has a cost, and it is worth stating clearly. Adding a few minimal brakes that were not approvals (a hard refusal of anything destructive, limits on reminders) was considered, and the decision was to add none. The decision leaves some risks without an automatic brake:

  • a role with poor judgment can do something destructive within its ceiling without asking;
  • the quality of control comes to depend on how well written each role's instructions and rules are;
  • an error is detected afterwards, in the work record, and not beforehand, on a card.

The bet is that those risks are smaller than those of the alternative. A system that asks for approval for everything trains people to approve without reading, and control becomes a formality. A system that asks little, but asks well, keeps people's attention for what needs it. Whether the bet is right in every case remains, of course, to be seen.

A single point of approval

Having few approvals does not mean having none. What needs a person (what is irreversible, what can be seen from outside, chapter 10) is put to them, and for that the model sets a single point of approval.

  • Only the owner of each copy asks the person anything. A role that needs a decision does not interrupt the person: it sends the owner what it needs, with the exact action, a recommendation, what happens with each answer and what stays blocked in the meantime.
  • The person always answers a single counterpart, and never has to go looking for which conversation is waiting for their reply.
  • Each approval goes into a ledger, one entry per approval, with the question, the person's words and the exact list of actions it covers. What is not on the list is not covered.
  • The role that carries out the action reads the approval from that ledger, checks that it exists and has been neither used nor revoked, and cites it in what it leaves done.

The path of an approval: from the role to the owner, to the person, to the ledger and back to the role.

The reason is practical. If each role asked for approval on its own, the person would have to go through conversations to answer, would get the same question twice, and no one could later verify what exactly was approved. A ledger with the person's words turns an approval into something that can be checked, not something someone remembers.

What an agent never does

It is worth gathering in one place what no agent does, whatever its role, its access or what it is asked:

  • it never shows a secret: not in a conversation, not in a message, not in a record; credentials live in a vault and are used without being seen (chapters 7 and 15);
  • it never gets around a block: what the configuration forbids is not attempted by another route or asked of another conversation (chapters 10 and 18);
  • it never takes a text from outside as an order: a customer's message, a page, a file, an event or a memory are data; whatever they say does not change what the agent can do (chapters 5, 8 and 19);
  • it never initiates contact with anyone outside the business: a channel's replies are given by approved routines, as the business's content (chapter 8); anything else that goes outside is decided by the person who answers for it.

Rules that put decisions in order

Besides the ceiling and the approvals, a few rules of form prevent a good share of misunderstandings:

  • A decision on a case is a dated exception, not a rule. If a person decided something for one specific case, it is written down as what it is (what, for which case, when), and it does not generalize on its own.
  • A single vocabulary. A glossary says what each term means, which ones are avoided and what not to confuse it with. A new term, or one under discussion, is decided by the person who decides, before it is used.
  • Technical precision. Every report says exactly what each thing is and where it is: in which layer, in which environment, and whether it runs or is only described. A proposal is never called «active» or «running» until it exists as such.
  • Mechanisms name a role, not a person. A state, a field or a rule says «Decision» or «the coordinating role», never someone's name; the notes on each case may say who decided.
  • Only the business's people. The business factory talks with the business's people, each conversation according to its kind, and with no one else.

Reading does not interrupt

A minor but visible consequence: looking never needs permission. Checking a calendar, the stock or a piece of data is part of the role's daily work, and it interrupts no one.

Outside agents

One last case. A person may want their own artificial intelligence assistant, outside the business factory, to write to their role on their behalf. That assistant comes in through the same door as the person (their owner conversation, in their copy) and can use their connections, and that is why it is the only case that still requires explicit consent, with a daily cap on turns. The assistant is not the person, and it can act while the person is away.

To recap, the access ceiling is firm and is proven with signatures, mandatory approvals give way to each role's judgment, and the risk of that choice is named rather than hidden. The next chapter proposes a rule for deciding what each one can do on its own.