Try using a subagent loop where you have one agent generate text and another critique the text according to some guidelines/goal.