As AI systems become more autonomous, the hardest problems are no longer about model quality - they’re about decision quality at scale. For years, AI systems optimized prediction: relevance, engagement, intent. Generative AI accelerated this approach. But the next shift is structural. We’re moving toward agentic systems that reason, plan, and act continuously in production. This keynote examines what actually changes when AI systems stop generating outputs and start making decisions. Drawing on experience operating large-scale personalization systems, we’ll explore how agentic architectures reshape production AI. Feedback loops tighten, failures become harder to detect, and optimization shifts from individual models to end-to-end decision systems. Small, specialized models are a critical enabler of this shift. Cheap to run and fast to iterate, they increasingly operate in coordinated swarms - handling routing, evaluation, policy, and memory. This is where cost, latency, and business leverage compound. As we head into 2026, advantage won’t come from the biggest model, but from teams that can design, govern, and operate decision-making systems responsibly at scale. Key takeaways: → Why agentic AI is a systems problem, not a tooling upgrade. → How small models create leverage in cost, latency, and control. → New failure modes introduced by self-reinforcing agent loops. → What to measure and guardrail when AI systems make decisions