Now in its third generation (AIRA₃), Meta FAIR's system pairs models with coding harnesses across many long-running, isolated agent instances that post hypotheses to a shared forum and pull solution artifacts from a shared filesystem, letting search strategies emerge from what individual agents choose to build on rather than top-down task assignment. In June 2026, Meta says an AIRA₃ ensemble placed 8th of ~4,000 teams for Gold in a live NVIDIA Kaggle competition to improve reasoning in a 30B Nemotron model. Prior versions are documented in AIRA₁ (arXiv:2507.02554) and AIRA₂ (arXiv:2603.26499).