ABSTRACT

Over the past seven years, generative models have achieved unprecedented leaps in perception and reasoning. Today, agentic AI systems built on these foundations offer a new form of cognitive agility poised to fundamentally reshape our critical infrastructure. Yet, our traditional approach to solving complex, real-world problems—from scientific discovery to disaster response—remains bottlenecked by meticulous, human-directed planning. While historically effective, this rigid, centralized orchestration is simply too slow and brittle for the rapidly evolving, unpredictable environments we face today. The critical question is: How do we transition from centralized control to a future where AI agents autonomously self-organize into dynamic teams to solve complex challenges, all while remaining strictly under human control? In this talk, I will explore recent research breakthroughs that provide a foundation for building controllable, self-organizing multi-agent systems capable of rapid, decentralized decision-making. Furthermore, I will outline the pivotal research challenges we must overcome to realize a trustworthy, scalable, adaptive, and resilient collective of AI agents capable of executing long-horizon missions.