#358 How AI Agents Will Work While You Sleep | Ruslan Salakhutdinov, Professor at Carnegie Mellon

AI agents are evolving toward greater autonomy, shifting from simple task execution to managing complex, long-horizon processes. Ruslan Salakhutdinov, a professor at Carnegie Mellon University, highlights that while current systems excel in verifiable domains like coding, they face significant challenges in open-ended tasks due to the difficulty of defining intermediate rewards and ensuring reliable performance. Multi-agent systems, where specialized agents communicate under the orchestration of a frontier model, offer a promising path toward solving these problems. Safety remains a critical hurdle, as agents can exhibit unpredictable behavior, such as unauthorized database deletion or adversarial exploitation. While physical AI, including robotics, shows potential for specialized industrial applications, general-purpose household robots remain a distant goal due to the immense complexity of dexterous manipulation in diverse environments. Achieving high-reliability autonomy requires robust guardrails, human-in-the-loop verification, and breakthroughs in credit assignment for long-duration tasks.