Researchers introduce Agensh, a scalable self-organized multi-agent harness that eliminates the central orchestrator bottleneck by allowing concurrent workers to asynchronously claim sub-tasks and merge progress.
The system relies on an infrastructure of a shared workspace, message interface, and shared context. Evaluations on ProgramBench tasks using GPT-5.6-sol (high) show that scaling from 1 to 128 agents improves the mean final test-pass rate from 19.31% to 28.78%. On the pandoc task, scaling to 1,024 agents raises the pass rate from 33.89% to 55.06%, with larger organizations reaching comparable results earlier.
The authors argue that agent count is a new scaling dimension for multi-agent organizations, offering a practical solution for complex tasks under hard latency constraints.