Agensh: Scaling Organizational Intelligence to 1,024 Agents

An orchestrator-free multi-agent harness scales to 1,024 self-organized agents, raising ProgramBench pass rates from 19.31% to 28.78% (1 to 128 agents on the hardest tasks) and from 33.89% to 55.06% on pandoc (1 to 1,024 agents).

Agensh is a self-organized multi-agent harness that removes the central orchestrator that typically bottlenecks multi-agent systems. Instead of a central coordinator, workers autonomously run a cooperation loop: they concurrently gather context, claim sub-tasks, act, share findings, verify results, and merge progress asynchronously, supported by a shared workspace, a message interface, and shared context. Evaluated with GPT-5.6-sol (high setting), on the five hardest ProgramBench tasks scaling from 1 to 128 agents raised the mean final test-pass rate from 19.31% to 28.78% (a relative improvement of roughly 49%), while on the pandoc task scaling from 1 to 1,024 agents raised the final test-pass rate from 33.89% to 55.06%. The paper appeared on Hugging Face daily papers on 2026-09-21 and on arXiv (cs.AI / cs.LG / cs.CL) on 2026-09-22. The two source reports agree on all facts; no conflicts were found.

• Agensh has no central orchestrator; workers asynchronously claim and self-assign sub-tasks.
• Cooperation is supported by shared infrastructure: a shared workspace, a message interface, and shared context.
• Evaluation model: GPT-5.6-sol, run at the high setting (per the arXiv report).
• Five hardest ProgramBench tasks: mean final test-pass rate rose from 19.31% (1 agent) to 28.78% (128 agents), roughly a 49% relative improvement.
• pandoc task: final test-pass rate rose from 33.89% (1 agent) to 55.06% (1,024 agents).
• Listed on Hugging Face daily papers on 2026-09-21 and on arXiv (cs.AI / cs.LG / cs.CL) on 2026-09-22.
• The two source reports agree on all figures; no conflicting facts were identified.