Self-organizing agent teams outperform oracle routers
Researchers from Stanford, Together AI, and Emory introduced Self-Organizing Agent Teams (SAT), a framework where autonomous LLM groups learn reusable coordination strategies without hand-engineered workflows. Evaluated across five math and physics benchmarks with o3-mini, Claude Sonnet 4, and DeepSeek-V3, the self-organizing team reached 66.7% accuracy, outperforming both the strongest single model (48.8%) and an oracle router (59.0%).
Hand-coded multi-agent graphs are rapidly becoming obsolete; granting agent teams the autonomy to self-organize and meta-optimize their own interaction protocols unlocks collaborative reasoning that fundamentally surpasses static orchestration and oracle model selection.
- –**Collaborative computation over routing:** The framework demonstrates that heterogeneous agents can collectively solve problems none could solve in isolation, moving beyond ensembling or selection into genuine multi-agent problem-solving.
- –**Remarkable sample efficiency:** Reusable communication and debate protocols were derived from as few as 15 training problems, proving transferable across distinct domains and competitions without problem-specific tuning.
- –**End of rigid topologies:** Static DAGs and pre-baked agent roles underperform adaptive, learned conversational dynamics that adapt to problem difficulty and model strengths.
DISCOVERED
1h ago
2026-09-23
PUBLISHED
2h ago
2026-09-23
RELEVANCE
AUTHOR
omarsar0