@albertvillanova we already had a discussion on this here: #526

Because passing no tool does not work with smolagents when we're trying to call tools. But maybe we could allow tool_choice="auto", and log an error if the LLM chooses to pass no tool?

@aymeric-roucher, thanks for the ping to the old PR: I had missed that one!

According to my understanding, for ToolCallingAgent, there's already a fallback mechanism for handling when chat_message.tool_calls is None. Specifically, we invoke self.model.parse_tool_calls(chat_message), which attempts to extract tool calls from chat_message.content.

• For example, even when I explicitly forced tool_choice="none", the default InferenceClient model was able to parse: chat_message.content = 'Action:\n{\n "name": "final_answer",\n "arguments": {"answer": "120"}\n}'.

Also note that if parsing fails and no tool calls are found, we already raise an error via:

Therefore, I think we are already covered in such cases through explicit parsing, checks and errors.