Running many models on Bedrock with `output_type` fails with "This model doesn't support the toolConfig.toolChoice.any field."

[shrik450]

Initial Checks

• I confirm that I'm using the latest version of Pydantic AI
• I confirm that I searched for my issue in https://github.com/pydantic/pydantic-ai/issues before opening this issue

Description

Running many bedrock models (I've tested Llama 4 and Nova) will always fail if the Agent has an output_type set. This happens even if there are NO tools in the agent. This seems to have been introduced in 0.2.12 and has persisted through to 0.3.4. The repro example I've included uses a pydantic Model, but this also happens with a dataclass or even a primitive like float. Versions 0.2.11 and before seem to be fine.

Example Code

#!/usr/bin/env -S uv run --script
# /// script
# requires-python = ">=3.13"
# dependencies = [
#     "pydantic-ai==0.3.4",
#     "dotenv",
# ]
# ///


from pydantic_ai import Agent


async def main():
    agent = Agent(
        model="bedrock:us.meta.llama4-scout-17b-instruct-v1:0",
        system_prompt="You are a helpful assistant.",
        output_type=float,
    )

    input_data = "How many pounds are in 10 kilograms?"

    response = await agent.run(input_data)

    print(response.output)


if __name__ == "__main__":
    import asyncio

    asyncio.run(main())

Output:

botocore.errorfactory.ValidationException: An error occurred (ValidationException) when calling the Converse operation: This model doesn't support the toolConfig.toolChoice.any field. Remove toolConfig.toolChoice.any and try again.

E: Simplified repro script.

Python, Pydantic AI & LLM client version

Python 3.13, pydantic-ai 0.3.4.

[DouweM]

@shrik450 Our Bedrock implementation lets you specify that the given model doesn't support tool choice:

agent = Agent(
        model="bedrock:us.meta.llama4-scout-17b-instruct-v1:0",
        system_prompt="You are a helpful assistant.",
        output_type=float,
        profile=BedrockModelProfile(bedrock_supports_tool_choice=False)
    )

Can you see if that works? If so, we can make this the default for Amazon and Meta models run through Bedrock, as we already do for Anthropic:

pydantic-ai/pydantic_ai_slim/pydantic_ai/providers/bedrock.py

Lines 57 to 68 in 81bd32b

	provider_to_profile: dict[str, Callable[[str], ModelProfile | None]] = {
	'anthropic': lambda model_name: BedrockModelProfile(bedrock_supports_tool_choice=False).update(
	anthropic_model_profile(model_name)
	),
	'mistral': lambda model_name: BedrockModelProfile(bedrock_tool_result_format='json').update(
	mistral_model_profile(model_name)
	),
	'cohere': cohere_model_profile,
	'amazon': amazon_model_profile,
	'meta': meta_model_profile,
	'deepseek': deepseek_model_profile,
	}

Feel free to file a PR to that effect!

[shrik450]

I tested both Amazon and Meta models with this fix and they both work! I'll take a look at cohere and deepseek as well just to make sure and have a PR out shortly.