Researchers at the University of California, Santa Barbara have introduced a new agent framework called Group-Evolving Agents (GEA), which matches or exceeds the performance of human-engineered AI systems while eliminating inference costs for deployment. This innovative approach addresses a key limitation in conventional AI systems, where agents are often confined to individual growth and can fail to share valuable insights, leading to lost innovations. By leveraging collective experiences within a group of agents, GEA allows for autonomous adaptation and self-repair, showcasing significant improvements in complex coding tasks on benchmarks like SWE-bench, where it achieved a 71.0% success rate compared to a baseline of 56.7%. This development suggests that organizations may eventually lessen their reliance on large teams of prompt engineers as agents become capable of independently optimizing their frameworks.
University of California, Santa Barbara develops Group-Evolving Agents framework for AI
