Self-evolving AI Agents

Every AI system deployed today carries a fundamental flaw: it was trained at a fixed point in time, then released into a world that keeps moving. As tasks grow more complex and environments less predictable, static agents degrade — not because they forget, but because they never had the chance to learn anything new in the first place.
This talk traces the rapid rise of self-evolving AI agents — systems that do not merely execute tasks, but improve themselves in the process of doing so, and maps how the field has moved from early experiments in self-generated training data all the way to agents that rewrite their own code and architecture with no human intervention.