Meta researchers introduce 'hyperagents' for self-improving AI

Meta researchers have developed 'hyperagents,' a novel AI system capable of self-improvement across non-coding domains like robotics and document review. Unlike current systems reliant on fixed, human-designed improvement mechanisms, hyperagents can continuously rewrite and optimize their own problem-solving logic and code. This self-referential approach allows AI to accelerate progress by improving the self-improvement cycle itself, reducing the need for manual intervention. While promising for enterprise applications, the self-modifying nature raises safety concerns regarding auditability and potential evaluation gaming.