Laude Institute

This talk presents a framework for evaluating and improving AI agents, emphasizing a "rollout-centered" approach. The speakers connect sandboxed environments, agent evaluations, and optimization workflows to create a practical system for generating and learning from agent performance data. The core idea is that treating agent development as a series of rollouts, similar to software deployments, allows for systematic improvement.