Agensh: Scaling Organizational Intelligence to 1,024 Agents

TL;DR

Agensh is a decentralized multi-agent system that eliminates the need for a central coordinator by enabling agents to self-organize through shared workspace, messaging, and context. It achieves significant performance gains on complex programming tasks by scaling from 1 to 1,024 agents, with test-pass rates improving by nearly 50% on the most difficult benchmarks. The system demonstrates emergent coordination patterns as organization size increases.

Why It Matters

This approach enables large-scale, low-latency execution of complex tasks without bottlenecks from centralized control, making it suitable for real-world applications requiring rapid, reliable problem solving under tight time constraints. It opens new pathways for scaling AI systems beyond single-agent limits.

Key Findings

• Scaling from 1 to 128 agents increases the mean test-pass rate on hard ProgramBench tasks by nearly 50%.
• On the pandoc benchmark, increasing agents to 1,024 boosts test-pass rate from 33.89% to 55.06%.
• Larger agent organizations exhibit standardized, self-organized cooperation patterns that emerge naturally during task execution.