Princeton paper finds teams of specialized AI agents beat single models

A Princeton paper finds specialized AI agents from many developers hit around 70% accuracy on SWE-bench, beating single general-purpose agents.

Published: 2026-10-20