What if the bottleneck isn't the model, it's how you're using it? The shift happening across serious AI engineering teams right now isn't about finding a better model. It's about moving from single-model architectures to compound systems where multiple models, tools, and retrieval components work together. The teams doing this are unlocking capability and reliability that no single frontier model can match on its own. This session breaks down how compound AI systems are being designed and deployed in production: where to split tasks across models, how to manage the interfaces between components, and what failure modes emerge when systems get more compositional. Key takeaways: - What compound AI systems are and where they outperform single-model approaches - How to design clean interfaces between models, tools, and retrieval layers - Where compositional architectures introduce new failure modes and how to catch them