Passing verbosity in extra_body breaks structured output

[holmbuar]

Sorry for not providing a reproducible example, but just for illustration, in my test code running the agent will obey the structured output fields, while agent2 will make up extra JSON fields in the final_output, triggering various Pydantic JSON validation errors. Am I using ModelSettings correctly?

openai                    1.100.2
openai-agents             0.2.8

from pydantic import BaseModel
from agents import Agent

class CalendarEvent(BaseModel):
    name: str
    date: str
    participants: list[str]

agent = Agent(
    name="Calendar extractor",
    instructions="Extract calendar events from text",
    output_type=CalendarEvent,
    model="gpt-5"
)

agent2 = Agent(
    name="Calendar extractor 2",
    instructions="Extract calendar events from text",
    output_type=CalendarEvent,
    model="gpt-5",
    model_settings=ModelSettings(
         reasoning={"effort":"low"},
         extra_body={"text":{"verbosity":"low"}}
     ),
)

[seratch]

Hi, thanks for asking the question. This is because you pass extra_body={"text":{"verbosity":"low"}}. The structured outputs use the "text" data structure, so passing the whole "text" object this way could break the things. Instead, you can pass Reasoning object like below. With this way, your verbosity parameter is safely merged into the text parameter including the structured output schema:

from agents import Agent, Runner, ModelSettings
from openai.types.shared import Reasoning
from pydantic import BaseModel

class CalendarEvent(BaseModel):
    name: str
    date: str
    participants: list[str]

agent2 = Agent(
    name="Calendar extractor 2",
    instructions="Extract calendar events from text",
    output_type=CalendarEvent,
    model="gpt-5",
    model_settings=ModelSettings(reasoning=Reasoning(effort="low"), verbosity="low"),
    # model_settings=ModelSettings(reasoning={"effort":"low"}, extra_body={"text":{"verbosity":"low"}}),
)