Support for function tool direct/structured output

[JRich3086]

Please read this first

• Have you read the docs?Agents SDK docs - Yes
• Have you searched for related issues? Others may have had similar requests - Yes

Describe the feature

Unless I've missed something, any agent tool output from final_output or new_item is forced to a string. I have a use case where it would be nice to be able to directly return an object from a function tool. For example, I am using Tavily search as a tool, which returns a dict, which I am converting to a Pydantic model. I know this can be done using an Agent as a tool and setting the output_type, but I'd prefer to just use the tool itself, if possible.

My workaround is as follows:

if new_item.type == 'tool_call_output_item':
    tav_json = eval(new_item.output)
    tav_response = PydanticModel.model_validate(tav_json)

But it would be nice to be able to convert the output directly in the function tool and return that:

@function_tool
def tav_search(query: str):
    """Search the web for information.
    
    Args:
        query: The user's query.
    """

    search_response =  TavilyClient.search(
        query=query,
        max_results=3,
        include_answer="advanced",
    )

    result = PydanticModel.model_validate(search_response)
    return result

Or possibly even set an output_type on the function tool itself.

[rm-openai]

Yeah this should work as you described. The output does need to be stratified before we send it to the LLM, so you should ensure your object has a good __str__ (or repr) but you can return anything you like from here.

[JRich3086]

Yeah this should work as you described. The output does need to be stratified before we send it to the LLM, so you should ensure your object has a good __str__ (or repr) but you can return anything you like from here.

Thank you. To clarify: this does work, but once the agent picks the tool result back up, the output is always forced to a string. And I don't want to strictly set a single output_type on this agent because it can also validly return strings when not calling this tool.

[rm-openai]

Thank you. To clarify: this does work, but once the agent picks the tool result back up, the output is always forced to a string. And I don't want to strictly set a single output_type on this agent because it can also validly return strings when not calling this tool.

Am i understanding correctly that you want the output of the tool to be the output of the agent? Maybe the agent's tool_use_behavior would be useful then, it lets you stop when a particular tool is called: https://openai.github.io/openai-agents-python/ref/agent/#agents.agent.Agent.tool_use_behavior

Let me know if there's a different desired behavior you want.

[JRich3086]

Thank you. To clarify: this does work, but once the agent picks the tool result back up, the output is always forced to a string. And I don't want to strictly set a single output_type on this agent because it can also validly return strings when not calling this tool.

Am i understanding correctly that you want the output of the tool to be the output of the agent? Maybe the agent's tool_use_behavior would be useful then, it lets you stop when a particular tool is called: https://openai.github.io/openai-agents-python/ref/agent/#agents.agent.Agent.tool_use_behavior

Let me know if there's a different desired behavior you want.

Thank you. I think tool_use_behavior should accomplish what I am looking for. I had not yet updated my agents module so I didn't see this option.