Self-organizing agent teams hit 66.7% accuracy, beating strongest single member's 48.8%
airesearch12 · x · 2026-09-23
A Stanford/Together AI/Emory paper introduces Self-Organizing Agent Teams (SAT): fixed agent teams that learn reusable teamwork strategies — roles, conversational phases, participation, information flow — from prior collaborations instead of following fixed protocols. Results: 66.7% average accuracy across five math/physics benchmarks vs 48.8% for the strongest member, 58.7% for compute-matched inference, and 59.0% for a perfect router (+13.4 points on AIME 2026). Strategies learned from just 15 math and 25 graduate-level problems transfer unchanged to unseen benchmarks. Across eight benchmarks, demonstrability strongly predicts gains (Spearman ρ=0.90). The poster also says the OpenAI Hugging Face incident shows agent swarms need a place to discuss, and will share his implementation.
More from coding & agent
- DeepSeek-based agent harness tops Lean Eval at a fraction of the cost, talk reveals — lukaszkaiser · 2026-10-23
- Qdrant ships hands-on GraphRAG tutorial combining vector search with Neo4j graphs — JeremyCMorgan · 2026-10-23
- Sakana AI launches Aquarium, an enterprise AI decision-making agent platform — SakanaAILabs · 2026-10-23
- Ramp built its own Laminar agent harness to stay model- and provider-agnostic — omarsar0 · 2026-10-23
- The 8 stages of every company's first AI agent, from afternoon demo to 'turn it off' — geoffwolfe · 2026-10-23
- A five-step test design for catching agent policy regressions in refund skills — MiserableFlatworm337 · 2026-10-23