Performance slowdown when enforcing `output_type` with a Pydantic `BaseModel` in an Agent

[MansourKama]

Question

Hi 👋

I’m using PydanticAI and defining an Agent with a Pydantic BaseModel as the output_type to ensure the model always returns a fixed structured output, even when hallucinations occur.

Here’s a simplified snippet of my setup:

module_agent = Agent(
    template_fallback_model,
    name=f"Agent Module {module_code}",
    output_type=list[pydantic_module],
    system_prompt=agent_system_prompt,
    tools=selected_tools,  # direct injection of matching tools
    retries=2,
)

Where pydantic_module is dynamically generated from a JSON template:

pydantic_module = build_pydantic_model_from_module(module_template)

def build_pydantic_model_from_module(module_template: dict) -> type[BaseModel]:
    ...

I noticed that when I enforce this output_type, the agent becomes significantly slower in returning responses.

My question:
👉 Is this slowdown expected because of internal schema validation or regeneration attempts when forcing the model to match a Pydantic schema?
If so, is there a recommended way to optimize structured outputs while keeping schema consistency?

Thanks in advance 🙏

Additional Context

No response

[DouweM]

@MansourKama I suggest adding https://ai.pydantic.dev/logfire/ to see where the time is lost; you can instrument both Pydantic AI itself, httpx for API requests, and anything else you're using including your own code that you can instrument yourself.

What you'll likely see is that the slowdown is coming from the API request itself, not from Pydantic Validation or Pydantic AI, meaning the model is taking more time to generate the output faithfully. You may have a better result with Native Output mode (https://ai.pydantic.dev/output/#output-modes) than the default Tool Output mode, which relies on tool_choice=required which has been shown to be slow with e.g. GPT-5.

You may also want to look at https://ai.pydantic.dev/output/#structured-dict as an alternative to build_pydantic_model_from_module.

[github-actions]

This issue is stale, and will be closed in 3 days if no reply is received.

[github-actions]

Closing this issue as it has been inactive for 10 days.