Self-Organizing Agent Teams paper: 15 practice problems, beats perfect router by 13.4 pts
rohanpaul_ai · x · 2026-09-25
The arXiv paper "Self-Organizing Agent Teams Learn to Reason Together" (2609.22682, Pappu, Suzgun, Zou et al.) shows fixed AI agent teams learning reusable teamwork strategies from just 15 math and 25 grad-level problems. Across five benchmarks teams average 66.7% vs. 48.8% for their strongest member and 59.0% for a perfect router, beating oracle routing by 13.4 points on AIME 2026. The paper also analyzes when demonstrability makes self-organizing collaboration pay off.
Related event: Stanford's self-organizing agent teams beat best single models(3 posts)→
More from coding & agent
- Inside Quail: custom vLLM scheduler, workload-aware KV cache for 1B tok/min — sh_reya · 2026-09-25
- Quail: open-source AI-SQL engine hits 1B+ input tokens/min on one H100 — sh_reya · 2026-09-25
- Anthropic's CI job volume grew 25x in six months — here's how they scaled test selection — JeremyCMorgan · 2026-09-25
- Perplexity's Fast Search powers Hermes Agent with 160ms p50 latency, free for all tiers — denisyarats · 2026-09-25
- Cursor launches Projects: one coordinator directing thousands of subagents, heavy users merge 6x more PRs — gaganghotra_ · 2026-09-25
- Paperclip lets any agent harness (Claude Code, Codex, Grok) work as employees in an AI company — AIFlow_ML · 2026-09-25