Researchers introduced Agensh, a multi-agent system without a central orchestrator. Workers self-assign tasks, share findings and verify results asynchronously. A shared workspace, message interface and context store support the loop. Scaling from 1 to 128 agents improved test-pass rates by 49% on ProgramBench. Scaling to 1,024 agents raised pass rates from 33.89% to 55.06% on pandoc. Larger groups reached targets sooner. Self-organized cooperation patterns emerged as organizations grew. The paper positions agent count as a new scaling dimension for general intelligence. It targets complex tasks under tight latency constraints. The work was submitted to arXiv on 22 September 2026.
What changed
Multi-agent systems relied on a central orchestrator to allocate tasks and coordinate workers.
What it unlocks
Running 1,024 agents without a central orchestrator to solve complex tasks faster.
- 1 to 128 agents: 19.31% to 28.78% test-pass rate
- 1 to 1,024 agents: 33.89% to 55.06% test-pass rate
- 49% relative improvement at 128 agents
- 5 ProgramBench tasks
What you need to act on it
- GPT-5.6-sol (high) model access
Sources