agent stop when no tool called

[CoderJzy]

Please read this first

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

Question

I want to only get the tool result by using Agent, but not the final response with further LLM processing.
And then i studied the docs deeply, I find there is a tool_use_behavior parameter for Agent which have a stop_on_first_tool attribute.
After using the stop_on_first_tool, I found it that agent continues to respond further LLM result when there is no tool called.
How can agent stop when either no tool is called or first tool called?

[seratch]

Thanks for asking the question. You may want to force running tools in the case. Please refer to https://openai.github.io/openai-agents-python/agents/#forcing-tool-use for more details. If you don't want to always run tools, there is no way to achieve your goal here.

[CoderJzy]

@seratch can i achieve the goal by creating a new tool which has no meaning and returns an empty string with the tool_choice="required" for Agent?
Then if agent doesn't want to use other tools, it will be required to use the empty tool to respond me an empty string.

[seratch]

I generally don't recommend the design, but if it works for your specific use case, it would be fine.

[CoderJzy]

@seratch I tried agent code as the following which equipped with three tools.

agent = Agent(
        name="agent",
        instructions="You are a helpful assistant",
        model=model,
        tool_use_behavior=StopAtTools(stop_at_tool_names=['compare', 'weather', 'nonsense']),
        tools=[weather, compare, nonsense],
        model_settings=ModelSettings(tool_choice='required')
)

But I found llm didn't always call a tool. For example when I say hello, it won't call any tool.
Is there any problem with my code? Or I confused the meaning about tool_choice="required"? Or It is due to my model's fault, I didn't use the openai model but my local ollama qwen model.