Domain concepts (Money, PhoneNumber, OrderId, Currency) represented as raw strings, numbers, or booleans — the type system can't tell two domains apart.
Function signatures use raw strings and numbers where domain concepts hide; the agent cannot tell from the type whether an argument is the right kind of thing.
Each domain concept has a small typed home — Money, PhoneNumber, OrderId — that knows its rules.
Each domain concept has its own typed wrapper; the agent's type checker catches wrong-primitive-in-wrong-slot before runtime.
Before the refactoring
function priceFor(cents, currency) {// ...}
After the refactoring
class Money { /* amount + currency, with arithmetic */ }function priceFor(money) {// ...}
Validation and formatting scatter across every consumer; the wrong primitive in the wrong slot compiles fine and fails at runtime; behavior that belongs with the concept has nowhere to live.
The agent must inspect call-site context (variable names, surrounding code) to verify a primitive is the right kind; validation and formatting scatter.
Wrapping every primitive is overkill — wrap when the concept needs validation, formatting, or domain-specific behavior beyond what the primitive offers.
Each wrapper is a class the agent must instantiate at every entry point; for primitives without domain rules the wrapper is overhead with no return.
Misuse becomes a type error; behavior accretes around the concept; refactoring is local to the wrapper.
Wrong-primitive misuse becomes a type error the agent catches without runtime testing; behavior accretes around the concept where the agent expects to find it.
Wrapping every primitive on principle — Name, Title, Description as separate classes whose only methods are toString() — adds typing ceremony without enforcing any domain rule.
Wrapping every primitive — including ones with no domain rules — adds boilerplate the agent must navigate at every signature with no reasoning benefit.