A boundary an agent cannot see, it walks straight through

Your system already has boundaries. They are just enforced by people remembering them. A review board that goes on recollection. An architect who knows which services you do not touch without asking. Real constraints, held for years because the people doing the work carried them in their heads.

An agent cannot do that. It cannot respect a boundary it cannot see, and it walks straight through the ones that live in a person’s judgement.

Writing the rule down is not enough. A rule in a document is still unenforced. Making a boundary explicit means turning it into something a machine checks. The word “policy” hides the work, because it does two jobs at once. Pull them apart and a boundary has four parts.

The constraint is the rule: customer data stays in the EU. The decision applies it, owned by a named architect: this service stores its data in Frankfurt. That decision spawns the invariant other systems rely on: this service’s data resides in Frankfurt. The check enforces it: the build fails the moment the service points at a store outside Frankfurt.

Now an agent can rework that service all day. The first time it reaches elsewhere, the build goes red.

This is what lets AI do the work. A boundary it can see, made of rules it cannot cross without tripping a test.