Everyone's shipping demos. Few are shipping systems. There's a brutal gap between a working prototype and AI that holds up in production and it lives in the parts no one talks about at conferences: data pipelines that drift, evaluation frameworks that don't catch what matters, and agent loops that fail in ways unit tests never predicted. Drawing from building Jules Google's autonomous engineering agent and post-training work at DeepMind, this session gets into the actual failure modes: where the architecture breaks, what monitoring doesn't catch, and the engineering decisions that separate systems that scale from ones that quietly degrade.