Agensh: Scaling Multi-Agent Collaboration to 1,024 Agents Without a Central Orchestrator
突破中心化瓶頸！Agensh 框架將多 Agent 協作無縫擴展至 1,024 個智慧體
Traditional multi-agent systems are constrained by a central orchestrator's coordination bottleneck. To resolve this, researchers introduced Agensh, a decentralized harness where concurrent agents run an asynchronous, self-organized loop of claiming tasks, sharing context, and merging progress. Evaluated with GPT-5.6-sol (high), scaling from 1 to 128 agents on ProgramBench improved the test-pass rate from 19.31% to 28.78%. On pandoc, scaling to 1,024 agents raised the pass rate from 33.89% to 55.06%, proving that scaling agent count is a viable path to enhanced performance.
Key points
Decentralized Self-Organization
Eliminates the central coordinator, utilizing an asynchronous self-assignment and progress-merging loop to prevent scaling bottlenecks.
Three Infrastructure Pillars
Relies on a shared workspace, a message interface, and shared context to sustain continuous and coordinated agent interaction.
Significant Scaling Performance
On the pandoc task, scaling the agent count from 1 to 1,024 boosted the final test-pass rate from 33.89% to 55.06%.
Emergent Cooperation Patterns
Analysis shows that distinct, standardized forms of self-organized cooperation spontaneously emerge as the agent organization grows.
How it works
Why it matters
This research establishes "agent organization scale" as a key scaling dimension for AI alongside parameter count and training data. By proving that massive asynchronous agent groups can resolve complex tasks under strict latency constraints, Agensh offers a practical paradigm for enterprise-grade, highly parallel AI workflows.
Who it affects
- AI Developer
- AI Researcher
- Enterprise Leader
- Product Manager
How to use it
- 1Large-scale software development and codebase refactoring
- 2Parallel processing of complex tasks under strict latency constraints
Limitations & caveats
- Self-organization without guidance may lead to redundant work or task conflicts
- Continuous context access by massive agents can incur substantial API and token costs
Related
Grow the Harness, Not the Context: Building Low-Cost Specialist Agents via Failure-Guided Code Synthesis
別撐大脈絡，改建構程式框架：Growing Harness 透過「失敗導向學習」自動生成高效能 Agent
Growing Harness is a novel paradigm that turns recurring agent control decisions into reusable, executable code via failure-guided synthesis, slashing inference costs by up to 98.6% while keeping small models highly capable.