Researchers at the University of California, Santa Barbara have developed Group-Evolving Agents (GEA), a framework in which groups of AI agents collectively share experiences and innovations to autonomously improve over time, rather than evolving in isolated lineages. Tested on coding benchmarks, GEA achieved a 71.0% success rate on SWE-bench Verified — matching the top human-designed open-source framework — and 88.3% on the Polyglot benchmark, compared to 68.3% for the prior state-of-the-art baseline. Crucially, evolved agents add no extra inference cost at deployment and retain performance gains when switched between underlying model providers such as Claude or GPT.

New agent framework matches human-engineered AI systems — and adds zero inference cost to deploy - VentureBeat