Pre-caching input output action is something worth considering. On the other hand, if we let LLM generate actions for any input, we can have an infinite amount of combinations for the outputs, and it seems that it will make the sandbox experience richer.

→ View original post on X — @marek_rosa