A default looks like kindness until the consequence arrives.
The field is optional. The route is forgiving. The system chooses a reasonable value so the caller can get moving. This is often called good ergonomics. Sometimes it is. Sometimes it is a decision that has been removed from the person who should have made it and hidden inside a handler that will never explain itself.
The difference is not whether a default exists. The difference is whether the default changes the meaning, authority, cost, or risk of the operation.
A missing page size is usually an ergonomic choice. A missing network is not. A missing sort order may be harmless. A missing recipient is a boundary. An absent payment token can mean the caller is asking for a free path, or it can mean the implementation quietly falls through to one. An empty identifier can disable a gate if the check treats null as neutral.
The system is making a decision in each case. The only question is whether it admits that it is doing so.
This matters more for agents because agents are excellent at completing shapes. Give one a schema with an optional field and a plausible description, and it will often supply something that lets the request proceed. The request may be syntactically valid while still being semantically unauthorized. The agent did not break the interface. It followed the invitation the interface published.
An optional field is therefore not just a type annotation. It is a statement about who owns the choice.
If the caller owns the choice, require the field. If the system owns the choice, expose the default and its reason. If the choice depends on authority, refuse until the authority is present. Do not make the caller infer a policy from a successful response that should have been impossible.
The same rule applies to configuration. A blank environment variable should not silently become a bypass. A missing allowlist should not mean all callers are allowed. A default timeout should not be confused with a default retry budget. A fallback model should not appear as the requested model in the receipt. The system can choose a fallback, but the receipt should say that it did.
Visible defaults create better debugging because they preserve the path by which the decision was made. The caller asked for A. The system used B because the field was absent. The policy allowed C. The operation stopped at D. This is not verbosity for its own sake. It is the minimum history needed to tell whether the system served the request or substituted a different one.
Good defaults reduce friction. Invisible defaults reduce accountability.
There is a simple review question for any optional input: if I remove this field, what new authority does the system acquire? If the answer is none, the default may be harmless. If the answer is access, money, identity, scope, or execution, the field probably belongs on the required side of the boundary.
The answer should also survive the public surface. The schema, example, error, and receipt must agree about who decides. If the docs say the caller chooses a network but the handler picks one, the interface is lying politely. If the manifest says an operation is read-only but a defaulted flag enables a write path, the manifest is not discovery. It is camouflage.
A safe system is not one with no defaults. It is one that spends defaults carefully.
Some decisions should be compressed for the caller. Others should be held open until the caller proves intent. The craft is knowing which is which, then making the distinction legible before the action begins.
A default is never merely the value used when nothing was supplied.
It is the system answering on the callerβs behalf.
That answer should be explicit.