Agents : connecting an outside tool
The connected agents page
An agent, here, is not a person: it is an open connection to a piece of software. Your ERP, your development workshop, your monitoring, or any tool you connect yourself, call your AI in your name, with nobody watching the screen at that moment.
Outside tools it can query, and what they bring.
Agents and colleagues are not the same thing. An AI colleague is someone you hire: it has a name, a trade, a tone, and it works for the people of the company from our screens. An agent is a door you open: it has neither name nor tone, it has a key.
How far it goes
Three scopes, each building on the last:
- Read, search your documents, nothing more;
- Chat, ask questions and receive answers;
- Act, chain tools together and propose entries.
An agent set to Read that is asked to chat receives a flat refusal, not a degraded answer. That is deliberate: a permission you can get around by insisting is not a permission.
Nothing is written without you
Even on Act, an agent only writes down what it would like to do. Every entry waits for a human hand, in the same Pending entries block as the AI colleagues', because a proposed entry is decided the same way whatever its origin.
The key, and the probe
At the end of the process a service key is minted: that is what you will paste into the tool's configuration. It is shown only once. The Probe button then answers "does this work?" before any real connection: it checks that the key opens, that the model is there and that the scope holds documents, without making the AI work, without consuming anything.
Revoking
Revoking closes the door immediately, wherever the key has been copied to. The agent stays in the list, marked as revoked, with everything it did. One does not remove from a file the trace of what was decided.
In closing
An agent calls your AI in your name, with nobody watching the screen. That is powerful, so it needs watching: leave connected only those you still use.
Commenting is not enabled on this course.