Publication
arXiv
Stage
Preprint
What we read
Summary of the paper
Authors
Zhihao Zhan, Ting Song, Li Dong, Shaohan Huang, Jianxun Lian, Yan Xia, Furu Wei
Universities and research institutions
Not yet supplied in verified metadata; the Brief does not guess.

What they did and found

In ProgramBench tasks, increasing agent count from 1 to 128 improved mean final pass rate by about 49% (19.31% to 28.78%), and extending to 1,024 agents raised pandoc’s final pass rate to 55.06% (from 33.89%). The gains accompany emergent self-organized cooperation patterns and no central orchestrator, with earlier highs at larger scales.

Why it matters

Shows scaling the number of agents as a distinct dimension for organizational intelligence, offering faster progress on hard tasks, but results depend on task type, budget, and implementation details.

Agensh runs a cooperation loop where workers gather context, claim sub-tasks, take action, verify results, and merge progress into a shared workspace. The infrastructure,shared workspace, message interface, and shared context,lets many workers contribute asynchronously, reducing bottlenecks.

Two-paragraph takeaway: more agents can improve both final results and speed on challenging tasks, but scale brings coordination challenges and requires robust workflows; benefits vary by task, budget, and model used, so teams should test incrementally.

Read the paper PDF ↗

Original sources · 1
  1. Agensh: Scaling Organizational Intelligence to 1,024 Agents ↗arXiv · 2026-09-22

Check the original paper for its authors, methods, version and access terms.

What remains uncertain

Limited to five tasks, fixed budget and model; results may not generalize to other tasks or settings.