Hello @taha-yassine ! Here switching to auto would introduce more of a problem than it fixes. smolagents required one tool call at every step: this exactly matches the argument required. Allowing for not having a tool call each time would break the lilbrary's logic. As such, inference solutions that do not support required argument, i.e. libraries that do not allow forcing a tool call, are simply incompatible with the logic of smolagents.

So let's close this: the only solution for your problem will be to modify the Model subclass itself to accept forcing a tool call!

Thanks for your answer. I see why auto could cause problems.
Do you think it would be worth it to refactor the Model class to use required by default but allow the user to override it?

@taha-yassine : overriding this argument to auto will not fix the issue either, because that will introduce the behviour where no tool is passed, which breaks the fundamental logic of this library. Thus it's better not to introduce the possibility to override.

I see. Looking closer at the code, the prompt suggests the agent use final_answer if it decides no tool needs to be called, which seems like a reasonable alternative. Combined with the ability to enable planning, I think it's good enough of a solution.
It's unfortunate that vLLM doesn't allow tool_choice="required". I will switch to another backend until it is implemented.
Thanks for you time!