What is a Structured Output?

By default, a language model produces prose, which is easy for humans to read but unreliable for software to parse programmatically — small wording variations can break a downstream parser. Structured output techniques (function/tool-calling schemas, JSON mode, grammar-constrained decoding) force the model's response to conform to a specific shape, so the calling code can deserialize it directly into typed objects with confidence the fields will be present and correctly typed. This is foundational to building reliable agentic systems: an agent deciding which tool to call, with what arguments, needs that decision expressed as structured data the orchestration layer can execute, not a paragraph describing the intent. Well-designed structured output schemas also improve the quality of the underlying reasoning, since forcing the model to commit to discrete fields (rather than a rambling explanation) often produces more precise, checkable answers.