AI can generate five implementations of the same rule before anyone notices the rule has split.
Generation speed makes duplicated truth more dangerous.
I have worked with systems where one business concept existed in an API schema, a database constraint, validation code, documentation and several tests. They matched when they were written. Months later, each represented a slightly different rule.
Agents amplify that drift because every context produces a plausible implementation. Contract-driven development offers a stronger boundary: humans define the semantics once, and tooling generates or validates the representations around it. The contract is not another document. It is an executable source of truth with versioning, ownership and compatibility rules.
AI can still help write the implementations. It should not be allowed to invent a new definition in each one.
When generation becomes cheaper, the source of truth has to become harder to duplicate.