Agensh: Scaling Organizational Intelligence to 1,024 Agents
Agensh reports higher ProgramBench pass rates by replacing a central orchestrator with self-organizing agent workers.
The paper says workers claim tasks, share findings, verify results, and merge progress asynchronously through shared workspace, messaging, and context layers. On five hard ProgramBench tasks with GPT-5.6-sol high, scaling from 1 to 128 agents raised mean final test-pass rate from 19.31% to 28.78%. On pandoc, 1,024 agents reached 55.06%, up from 33.89% with one agent. The authors frame agent count as a scaling dimension for complex tasks under tight latency or time budgets. ArXiv · AI/CL/LG's note
The paper says workers claim tasks, share findings, verify results, and merge progress asynchronously through shared workspace, messaging, and context layers. On five hard ProgramBench tasks with GPT-5.6-sol high, scaling from 1 to 128 agents raised mean final test-pass rate from 19.31% to 28.78%. On pandoc, 1,024 agents reached 55.06%, up from 33.89% with one agent. The authors frame agent count as a scaling dimension for complex tasks under tight latency or time budgets. ArXiv · AI/CL/LG's note
score 6