A connection is the way an agent reaches a system: email, the calendar, the management system, a spreadsheet. Without connections, an agent can only talk; with them, it can look at real data and act.
The role's and the person's
Connections come in two kinds, and the difference between them is one of the central decisions of the model:
- a role connection belongs to the role, and is used by its main conversation, its routines and its tasks;
- a personal connection belongs to a person (their email, their account on a service), and is used only in that person's copy, and only when they ask for something that needs it.
The rule fits in one sentence: the role's connections are for the whole role; a person's are used only in their copy and when they ask. Behind it is a fairly intuitive idea of ownership. A person's email is theirs: there is no reason for the whole area to read it, or for a routine that runs at night to go into that mailbox. What belongs to the role is used by the role; what belongs to each person is used only by that person, when they ask.
Connecting something of one's own, moreover, is always done by the person, never by an agent: an agent can point out where it is done, but the act of connecting, changing or deleting a connection belongs to the person.
Credentials never reach the agent
The access keys to each system do not live in the database of the business definition or on the business factory's machine. They live in an encrypted vault, with a space of its own for each business, from which they cannot be read back: they can only be used to call the service. A person's connection to an outside service, in turn, is made in the person's own browser, through the service's own flow, so that the system never sees their password.
The goal is that if someone one day gains access to the database or to the business factory's machine, they will not find a single key. And there is a second goal, specific to agent systems: an agent that never sees a credential cannot leak it, whether by mistake or because a text from outside persuaded it to.
A single door
Every call the agents make to the business's systems goes through a single gateway. Each conversation receives an address of its own that shows it only what concerns it (its role's connections and, if the person asked, that person's), and it is the gateway that calls the service with the right credential.
The reason for concentrating everything in one door is operational. With a single door, one can see in one place who called what, and cut it off in one place if something goes wrong. Organizations that connect hundreds of agents to hundreds of systems tend to reach the same conclusion: a central gateway, not one cable per agent.
Limits within each connection
Each connection can also restrict which operations the agent may use (a list of what is allowed) and fix certain values so that the model does not make them up.
Example. At Ópticas Horizonte, Sales has a connection to the stores' appointment system. It can only look up appointments and create one; it cannot delete. The store's identifier is fixed in the connection, so the agent cannot get the branch wrong.
What has no way to connect
Some systems offer no way to connect through software. For those cases, the model uses a remote browser: the person logs in once, by hand, in a live view that only they can see, and the agent then works with that login, browsing with a copy that does not keep changes. It is a compromise (slower and more fragile than a direct connection), but it covers what would otherwise be left out.
Outfitting a connection
Connecting a system is not the same as knowing how to use it. For the most important systems, especially those that can only be reached through a browser, an intermediate step is worthwhile: test the connection thoroughly first, and outfit it afterwards.
The test goes through the system for real: what can be read, where each piece of data is, what fails. With what was learned, a loadout is put together for the role:
- read tools, each one validated against the real system before it is accepted;
- proposed write tools, which stay as proposals until a real business case that needs them comes up;
- skills that explain how to use that system and how this business uses it;
- routines that read what matters and give notice only when something has really changed.
The asymmetry between reading and writing is deliberate. Reading badly produces a wrong piece of data that can be corrected; writing badly can change something in an outside system. First one explores and validates, and only then builds on solid ground.
One tool, one guide
Each connection can be linked to the skill that explains how to use that system: its map, its common tasks, its pitfalls. Connection and guide go together, because an agent that can get into a system but does not know how to move around in it tends to do more harm than one that cannot get in.
To recap, the role's connections belong to the role and each person's belong only to that person; credentials live beyond the agents' reach; and everything goes through a single door. The next chapter looks at the work that runs without anyone asking for it.