Structured Output With Output Type

The problem with hoping for JSON

In LangChain, getting structured data out of a model usually meant either careful prompt engineering, "respond only with JSON matching this schema," or with_structured_output(SomeModel), which works, but is bolted onto a framework that otherwise treats output as text.

Pydantic AI makes this the default way of thinking about an agent: you pass a Pydantic BaseModel as output_type, and run_sync either returns a validated instance of it or raises. There's no separate "structured mode," it's the same Agent and the same run_sync.

Declaring an output type

from dotenv import load_dotenv
from pydantic import BaseModel
from pydantic_ai import Agent

load_dotenv()


class CityFact(BaseModel):
    city: str
    country: str
    population_millions: float


agent = Agent("google:gemini-3.5-flash-lite", output_type=CityFact)


def main() -> None:
    result = agent.run_sync("Tell me about Paris.")

    fact = result.output
    print("Output type:", type(fact))
    print("City:", fact.city)
    print("Country:", fact.country)
    print("Population (millions):", fact.population_millions)

Under the hood, Pydantic AI turns CityFact into a JSON Schema, gives it to the model as the required output shape, and validates whatever comes back against the BaseModel before handing it to you. If the model returns something that doesn't validate, Pydantic AI feeds the validation error back to the model and asks it to try again, up to retries, default 1, you don't have to write that retry loop yourself.

Why this matters more than it looks

Every field on result.output is a real, typed Python attribute. Your editor autocompletes it, your type checker catches typos in it, and downstream code, saving to a database, rendering a UI, never has to defensively check if "population" in parsed_json. This is the single biggest practical difference from LangChain's default string output.