Meta’s AIRA₃ turns parallel agents into a research team
The interesting part is not that Meta built another agent. AIRA₃ placed eighth out of roughly 4,000 teams in an NVIDIA Kaggle contest, fine-tuning a 30B Nemotron model against a private test set. It gets there with long-running agents working in isolated environments, sharing hypotheses through a forum and artifacts through a filesystem, without a central planner. That architecture points toward research systems that improve by coordinating exploration, not just by scaling one model.