What is structured output?

Structured output is a language model responding in a fixed, machine-readable format, such as a set of named fields with defined types, rather than free text. Instead of "I'd suggest refunding about 40", the model returns a record with an action, an amount and a reason. For AI agents it is what lets other systems read, check and act on the model's output reliably.

Why does structured output matter once agents act?

Because actions need precise inputs. A payment system cannot act on a paragraph. A case system needs a category from its own list, not the model's paraphrase of one. Structured output gives every value a named place, which means every value can be checked.

That checking is the real gain. When the amount is a field, you can compare it with the transaction. When the account is a field, you can confirm it exists and belongs to this customer. Free text hides those values in sentences; structure exposes them.

How does structured output work in practice?

The developer defines a schema: the fields, their types, which are required and which values are allowed. The model is instructed, or constrained by the provider's tools, to return output matching that schema. The application validates the response. Anything that fails validation is rejected or retried, never guessed at.

Good schemas for agents share some features:

• Enumerated values for anything with a fixed list, such as action types or complaint categories.
• A source reference for every factual value, pointing to the record it came from.
• A separate field for the agent's explanation, so prose never leaks into values.
• An explicit "cannot determine" option, so the model is not forced to fill a field it has no basis for.

What does that look like in an example?

A freight forwarder's agent reads incoming customs queries and prepares responses. Its schema requires a shipment reference, a query type from a fixed list, the documents needed, and a draft reply. On one email the model returns a shipment reference in the right format that does not exist in the shipment system. The schema accepted it, because it was well formed. The next check, looking it up, did not. The case went to a person.

Is valid structure the same as a correct answer?

No, and this is the most common misunderstanding. A schema guarantees shape: the right fields, the right types. It says nothing about whether the values are true. A model can return a perfectly valid record containing an invented account number. Structure makes hallucination easier to catch, because each value can be checked, but it does not prevent it. Validation has two layers: does it fit the schema, and does it match the records.

How is structured output different from function calling?

Function calling is a specific use of structured output: the model returns a structured request to run a named function with arguments. Structured output is wider. It covers any response in a fixed format, including classifications, extractions and reports that never trigger a function. Both rely on the same idea, that a large language model should hand machines data, not prose.

Where Pryme fits

Pryme checks structured actions field by field against the agent's bounds and the provenance of the facts each value rests on. It reads from your existing systems rather than replacing them; see does an agentic operating system replace your existing systems? and the governed agentic operating system, explained.