Agensh README

项目的 README,来自 github.com/microsoft/Agensh。合并为一个 markdown 文件:README.md。

Agensh: A Decentralized, Scalable Multiagent Harness

Project Page Paper PDF arXiv Report X Post

News

⚙️ 2026-10-29: Agensh now supports Copilot, Codex, Claude Code, and Pi as single-agent backends.

🚀 2026-10-24: We open-sourced Agensh, including the harness, experiment launch scripts, and reproduction guide.

📢 2026-09-22: We published the Agensh preprint on arXiv.

Overview

Agensh is a self-organized multi-agent harness that scales to 1,024 agents. Agents discover and claim sub-tasks, share findings, verify results and merge progress asynchronously, without a central orchestrator. A shared workspace, message interface and shared context connect each agent's local progress to the team.

We evaluated Agensh on the five hardest ProgramBench tasks using GPT-5.6-sol with high reasoning effort, a six-hour budget and no Internet access. Agents rebuild reference programs from behavioral observations and bundled documentation. Across the 5 evaluated tasks, the mean final test-pass rate rises from 19.31% with 1 agent to 28.78% with 128 agents, an improvement of approximately 49% relative. Larger teams also reach comparable pass rates sooner. On Pandoc, the final test-pass rate rises from 33.89% with 1 agent to 50.94% with 128 agents and 55.06% with 1,024 agents under the same six-hour budget. We find that self-organized peer coordination, integration workflows and specialized roles gradually emerge as the team grows.

With GPT-6.1-sol at high reasoning effort, the six-hour Pandoc test-pass rate rises from 56.19% with 1 agent to 79.59% with 32 agents, a relative improvement of 41.64%.

Pandoc scaling results for GPT-5.6-sol and GPT-6.1-sol at high reasoning effort, with emergent coordination workflows and specialized roles

Watch 1,024 agents work together

Animated replay of 1,024 agents cooperating through Gitea, Mattermost and shared context

How it works

Agensh cooperation loop: gather context, claim a sub-task, take action, verify results and merge progress asynchronously through shared workspace, messaging and context

Agensh defines the multi-agent cooperation loop. Workers concurrently gather context, claim sub-tasks, take action, verify results and merge progress in a self-organized manner, repeating until the user’s goal is achieved. All workers proceed asynchronously and share progress through the infrastructure. Gitea provides the shared workspace, Mattermost provides the messaging interface, and the shared context mechanism is adapted from DeLM. In our 1.0 experiments, we integrate Agensh on top of the per-agent agentic loop of Copilot harness. We also support Pi, Codex, and Claude Code as underlying single-agent harnesses.

Reproducing the experiments

This release includes a single-node team setup using Docker Compose and a standalone single-agent baseline. Copilot is the default. Select Pi, Codex, or Claude Code with --agent-harness pi, --agent-harness codex, or --agent-harness claude when preparing and launching runs. Pass --harness-version VERSION to choose a harness release. Each worker image includes only the selected harness. All expose ACP to the existing infrastructure; Pi, Codex, and Claude Code use adapters to their native runtimes. Copilot is the harness used for the 1.0 experiments.

Citation

@article{zhan2026agensh,
  title   = {Agensh: Scaling Organizational Intelligence to 1,024 Agents},
  author  = {Zhan, Zhihao and Song, Ting and Dong, Li and Huang, Shaohan and Lian, Jianxun and Xia, Yan and Wei, Furu},
  journal = {arXiv preprint arXiv:2609.26781},
  year    = {2026},
  url     = {https://arxiv.org/abs/2609.26781}
}

License and contributing

Code is licensed under the MIT License. See the contribution guidelines, security policy, support information, and trademark guidance.

Agensh (microsoft/Agensh)

正在搜索数据库…

926 ·GitHub
这是什么?Agensh 是微软开发的一个自组织、去中心化的多代理框架,最多可让 1,024 个 AI 代理在没有任何中央协调器的情况下协同完成一项任务。
有没有中央协调器来管理这些代理?没有中央控制器。代理各自独立地发现并认领子任务、分享发现的结果、互相验证对方的成果,并通过共享工作区和消息层异步合并进展。
我该如何让它运行起来?发布版本提供了基于 Docker Compose 的单节点团队环境,以及一个独立的单代理基线版本;按照复现指南构建工作节点镜像并启动运行即可。
它支持哪些编程代理后端?Copilot 是默认的底层单代理框架,但启动运行时可以通过 --agent-harness 参数改选 Pi、Codex 或 Claude Code。
运行其基准测试需要联网吗?在已发布的评测中,代理在六小时的预算内工作,且不能访问互联网,只能依据行为观察结果和随附文档重建参考程序。
是什么基础设施支撑着共享工作区和代理之间的通信?Agensh 使用 Gitea 作为共享工作区,使用 Mattermost 作为代理之间的消息通道,并采用了改编自 DeLM 项目的共享上下文机制。
增加更多代理能持续提升结果吗?在一定范围内可以。在五个 ProgramBench 任务上,平均测试通过率从 1 个代理时的 19.31% 提升到 128 个代理时的 28.78%(相对提升约 49%),而且更大的团队收敛得也更快。
这背后有研究论文吗?有。Agensh 的设计与实验已写入一篇预印本论文,于 2026 年 9 月发布在 arXiv 上。
评估它使用了什么基准测试?Agensh 在五个最难的 ProgramBench 任务上进行了测试。这些任务要求代理依据行为观察结果和随附文档重建参考程序,测试中使用了 GPT-5.6-sol 和 GPT-6.1-sol 等模型。
作者
microsoft
大小
37 MB
星标
926
时长
2个月
提交 · 90 天