Found first: a primary source the press has not covered yet.
A Stanford-led team reports that groups of AI agents can learn collaboration strategies that outperform not just their strongest individual member, but also a perfect oracle that always selects the best member for each problem. Their paper, Self-Organizing Agent Teams Learn to Reason Together, records 66.7% average accuracy across five mathematics and physics benchmarks, against 59.0% for the perfect router and 48.8% for the strongest individual agent.
What the source says
Aneesh Pappu, Mirac Suzgun, and co-authors including Stanford's Mykel Kochenderfer and James Zou trained fixed teams of agents on 15 mathematics and 25 graduate-level knowledge problems. The teams learn reusable strategies that govern roles, conversational phases, participation, and information flow, and those strategies transfer without modification to unseen benchmarks. On AIME 2026 specifically, team performance exceeds the perfect router by 13.4 points. The authors identify a property they call demonstrability, the ability to distinguish correct from incorrect reasoning, which correlates with collaboration gains at Spearman ρ=0.90 (p=0.005) across eight benchmarks.
Why it matters
The oracle router is the ceiling for any selection-based approach: it always picks the single best agent per problem. Beating that ceiling means the gains come from team reasoning producing solutions no individual member reaches alone. The strategies generalize from 40 training problems to held-out benchmarks, which means the learned collaboration structure is not overfit to a specific domain. For practitioners designing multi-agent systems, this result shifts the design question from which model to pick toward how agents should be structured to reason together.