Every team building generative AI hits the same fork eventually: keep pushing one model to do everything, or start splitting the work across specialised components. Neither answer is obviously right, and the teams getting it wrong are finding out the expensive way. This session brings together practitioners who have landed on different sides of that decision, covering when compound architectures actually improve reliability and when they just add complexity and cost without a real payoff. Expect disagreement on where the line sits. What this session will cover: - When splitting a system into specialised components improves reliability, and when it doesn't - How teams are deciding between one capable model and several coordinated ones - The hidden costs of compound systems that don't show up until production - Real tradeoffs teams have made, including ones they'd reverse