I did develop my own agent around MiniMax. I did see weird behavior when I messed up the loop, like omitting pieces of remove thinking; maybe it's an agent bug, some models/providers just ignore/normalize the broken input, some don't.