[{"data":1,"prerenderedAt":63},["Reactive",2],{"news-story-news-20260924-agensh-1024-agents-swarms":3},{"id":4,"title_zh":5,"title_en":6,"summary_zh":7,"summary_en":8,"takeaways_zh":9,"takeaways_en":15,"content_zh":21,"content_en":22,"external_links":23,"source":36,"category":43,"impact_score":44,"views_count":45,"likes_count":46,"related_entities":47,"tags":55,"image_url":62},"news-20260924-agensh-1024-agents-swarms","微软研究院开源 Agensh：无中心调度的多智能体编码框架，最多 1,024 个 Agent 协作","Microsoft Research open-sources Agensh, a self-organizing harness that scales to 1,024 coding agents","微软亚洲研究院（MSRA）研究员、论文共同一作 Zhihao Zhan 于 10 月 24 日在 X 上宣布开源 Agensh（MIT 协议）。它是一套没有中心调度 Agent 的多智能体编码框架，各 Worker 自行认领子任务、验证并异步合并进度；官方口径下，在 pandoc 任务上从 1 个扩到 1,024 个 Agent，最终测试通过率由 33.89% 升至 55.06%。","Zhihao Zhan, a Microsoft Research Asia researcher and co-first author, announced on X on Oct 24 that Agensh is now open source under MIT. It is a multi-agent coding harness with no central orchestrator: workers claim sub-tasks, verify and merge asynchronously. Per the authors, scaling from 1 to 1,024 agents lifts the final pandoc test-pass rate from 33.89% to 55.06%.",[10,11,12,13,14],"10 月 24 日开源 harness、实验启动脚本与复现指南，MIT 协议，代码见 [microsoft/Agensh](https://github.com/microsoft/Agensh)。","官方口径：ProgramBench 最难 5 题、GPT-5.6-sol（high）、6 小时预算、断网条件下，平均最终测试通过率由 1 个 Agent 的 19.31% 升到 128 个的 28.78%，相对提升约 49%。","官方口径：pandoc 任务 1 → 128 → 1,024 个 Agent，通过率 33.89% → 50.94% → 55.06%。","协作基础设施：Gitea 做共享工作区、Mattermost 做消息、共享上下文机制改编自 DeLM；每个 Agent 的单体循环基于 Copilot harness。","本次发布包含 Docker Compose 单节点团队环境和单 Agent 基线，均使用原生 Copilot ACP 会话。",[16,17,18,19,20],"Open-sourced on Oct 24 under MIT: harness, experiment launch scripts and reproduction guide at [microsoft/Agensh](https://github.com/microsoft/Agensh).","Authors report: on the 5 hardest ProgramBench tasks (GPT-5.6-sol high, 6h budget, no Internet), mean final test-pass rate rises from 19.31% with 1 agent to 28.78% with 128 agents, about 49% relative.","Authors report: on pandoc, 1 to 128 to 1,024 agents moves the pass rate from 33.89% to 50.94% to 55.06%.","Coordination runs through Gitea (shared workspace), Mattermost (messaging) and a shared-context board adapted from DeLM; each worker uses the Copilot harness agent loop.","The release ships a single-node Docker Compose team setup plus a single-agent baseline, both on native Copilot ACP sessions.","### 核心背景与行业痛点\n\n现有多智能体编码框架大多依赖一个中心调度 Agent 来分配任务、协调 Worker，调度者的处理能力很快成为瓶颈，团队规模一大就难以继续扩展。微软研究院团队 9 月 22 日在 arXiv 发布 Agensh 预印本（2609.26781），10 月 24 日正式开源代码。本条为本站 9 月旧稿的更新版：此前版本中的部分机制描述和数字无法在论文与仓库中找到出处，现已按官方仓库与论文改写。\n\n### 架构亮点与底层机制\n\nAgensh 去掉了中心调度者。所有 Worker 运行同一个协作循环：收集上下文、认领并自我分配子任务、执行、验证结果、异步合并进度，直到目标完成。支撑循环的是三件基础设施：Gitea 托管共享代码仓库，记录待办、进行中与已完成工作；Mattermost 作为消息接口；共享上下文保存可复用的发现和工作意图（改编自 DeLM）。论文观察到，随着团队变大，同伴协调、集成流程和专职角色会自发出现并逐渐固定。\n\n### 权威 Benchmark 与实测跑分对比\n\n以下均为论文与仓库的官方口径，暂无第三方复现：在 ProgramBench 最难的 5 个任务上（让 Agent 依据行为观察和随附文档从零重建参考程序，GPT-5.6-sol high、6 小时、断网），平均最终测试通过率从 1 个 Agent 的 19.31% 升到 128 个的 28.78%；pandoc 单项从 33.89% 升到 128 个的 50.94% 和 1,024 个的 55.06%。论文同时指出，团队越大越早达到相同通过率。\n\n### 开发者实战落地与开箱指南\n\n仓库提供复现指南、团队启动器和单 Agent 启动器，可用 Docker Compose 在单节点拉起整套环境，并与单 Agent 基线对比。需要注意，实验依赖 Copilot ACP 会话与 GPT-5.6-sol，大规模并发的算力和 API 成本官方未公布，建议先用小团队规模验证。","### Core Background & Industry Pain Points\n\nMost multi-agent coding harnesses rely on a central orchestrator to assign work, and that orchestrator caps how far a team can scale. Microsoft Research posted the Agensh preprint (arXiv 2609.26781) on Sep 22 and open-sourced the code on Oct 24. This is an update of our September story: some mechanisms and figures in the earlier version could not be traced to the paper or repo and have been rewritten from official sources.\n\n### Architecture Highlights & Internals\n\nAgensh has no central orchestrator. Every worker runs the same loop: gather context, claim and self-assign a sub-task, act, verify, and merge progress asynchronously until the goal is met. Three pieces of infrastructure support it: Gitea as the shared workspace for proposed, ongoing and completed work; Mattermost as the message interface; and shared context for reusable findings and work intentions, adapted from DeLM. The paper reports that peer coordination, integration workflows and specialized roles emerge as the team grows.\n\n### Authoritative Benchmarks & Measured Scores\n\nAll figures are author-reported with no third-party reproduction yet. On the 5 hardest ProgramBench tasks (rebuild reference programs from behavior and bundled docs; GPT-5.6-sol high, 6h, no Internet), the mean final test-pass rate rises from 19.31% with 1 agent to 28.78% with 128 agents. On pandoc it climbs from 33.89% to 50.94% (128 agents) and 55.06% (1,024 agents). Larger teams also reach comparable pass rates sooner.\n\n### Developer Hands-on Guide\n\nThe repo includes a reproduction guide, a team launcher and a single-agent launcher; a Docker Compose single-node setup lets you compare a team against the single-agent baseline. Experiments depend on Copilot ACP sessions and GPT-5.6-sol, and compute or API cost at large scale is not disclosed, so start with small team sizes.",[24,28,32],{"title":25,"url":26,"type":27},"microsoft/Agensh (GitHub, MIT)","https://github.com/microsoft/Agensh","github",{"title":29,"url":30,"type":31},"Agensh project page","https://agens-harness.github.io/project/","demo",{"title":33,"url":34,"type":35},"arXiv 2609.26781: Agensh: Scaling Organizational Intelligence to 1,024 Agents","https://arxiv.org/abs/2609.26781","arxiv",{"platform":37,"author_name":38,"author_handle":39,"author_avatar":40,"tweet_url":41,"published_at":42},"twitter","Zhihao Zhan (Microsoft Research Asia)","@zhan8855","https://unavatar.io/x/zhan8855","https://x.com/zhan8855/status/2113974676661932342","2026-10-24T12:43:33.000Z","tooling",88,233,0,{"models":48,"agents":50,"vendors":53},[49],"gpt-5.6-sol",[51,52],"agensh","github-copilot",[54],"microsoft",[56,57,58,59,60,61],"Multi-Agent","Agensh","Microsoft Research","Open Source","ProgramBench","Coding Agents","https://agens-harness.github.io/project/figures/scaling.svg",1793718873817]