Every agent on its own
It worked with two. With twenty, each had its own credentials, every change broke an unknown number of them, and there was nowhere to look at what the whole thing had done.
Principles
Not brand values. Design decisions taken once, paid for properly, and not reopened on every project. A system is what it does when nobody is watching, and that is set by its rules.
Written after breaking them
The rules
It has its own coordinate, a profile and a trail. Without that you cannot answer who did what, and a system that cannot answer that is not auditable no matter how much it logs.
Memory, a trail, and a memory of that trail. If nothing is left when the session ends, tomorrow it starts just as clueless as today.
Across sessions, across computers, across containers that get rebuilt. If it depends on someone's laptop, it is not memory: it is luck.
If a capability is not behind it, it does not get improvised outside — it gets brought in. Today's exception is next year's invisible piece.
No general-purpose agent. One that does everything answers for nothing, and its context ends up too big to fit in anyone's head.
The credential is one and identity travels separately. That way rotating the key does not mean touching twenty places, and knowing who is speaking does not depend on which key they carry.
Where they come from
None of them was written before it was needed. They are here because one day something broke in a specific way and we did not want to go through that again.
It worked with two. With twenty, each had its own credentials, every change broke an unknown number of them, and there was nowhere to look at what the whole thing had done.
The cost did not disappear: it was paid once, in the right place. That is the difference between paying up front and paying every time something fails.
Changing them
It takes a real situation where the rule gets in the way and the specific damage it does can be shown.
What stops working, who it affects and what is lost. A rule was put there for a reason: removing it gives back the problem it solved.
Rules are changed by a human and there is a record of who and when. If an agent could rewrite its own rules, they would stop being rules.