Use structured outputs

Forcing the model to return validated, schema-constrained data removes an entire class of parsing and formatting failures before they can happen.

Why free text is fragile

If your agent returns prose that your code then parses, every small change in wording can break the parse. Structured outputs flip this around: you define a schema, and the model is constrained to fill it. The response is either valid against the schema or rejected, so malformed output never reaches your logic.

How to do it

• Define the shape you want with a schema. Structured Outputs guide and equivalents enforce it at decode time.
• Validate the result against a model such as Pydantic, and reject then retry on anything invalid.
• Libraries like Instructor wire this loop together so you get typed objects back instead of strings.

pythonfrom pydantic import BaseModel, ValidationError

class Refund(BaseModel):
    order_id: str
    amount_cents: int
    reason: str

def parse_refund(raw_json):
    # Return a typed object, or None so the caller retries instead of
    # acting on malformed output.
    try:
        return Refund.model_validate_json(raw_json)
    except ValidationError:
        return None

Validate, do not trust

Schema-valid is not the same as correct. A field can be well-formed and still wrong. Structured outputs remove the format failures so you can spend your reliability budget on the failures that actually require judgment.