Agensh README

README проекта с github.com/microsoft/Agensh. Одним файлом markdown: 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)

Ищем в базе…

926 ·GitHub
Что это такое?Agensh — это самоорганизующийся децентрализованный харнесс для работы нескольких агентов, созданный Microsoft. Он позволяет до 1024 агентам ИИ координировать работу над задачей без какого-либо центрального оркестратора.
Есть ли центральный оркестратор, который управляет агентами?Нет, центрального контроллера нет: агенты самостоятельно находят подзадачи и берут их на себя, делятся результатами, проверяют работу друг друга и асинхронно объединяют прогресс через общее рабочее пространство и систему обмена сообщениями.
Как это запустить?Релиз включает командную настройку на одном узле через Docker Compose, а также отдельную базовую версию с одним агентом. Чтобы собрать образы рабочих агентов и запустить расчёт, следуйте руководству по воспроизведению.
Какие бэкенды агентов для программирования поддерживаются?По умолчанию используется Copilot как базовый харнесс одного агента, но при запуске можно выбрать Pi, Codex или Claude Code с помощью флага --agent-harness.
Нужен ли интернет для запуска его бенчмарков?В опубликованных оценках агенты работают в рамках бюджета в шесть часов без доступа в интернет и восстанавливают эталонные программы исключительно по наблюдениям за их поведением и прилагаемой документации.
Какая инфраструктура обеспечивает общее рабочее пространство и обмен сообщениями между агентами?Agensh использует Gitea для общего рабочего пространства и Mattermost для обмена сообщениями между агентами, а также механизм общего контекста, адаптированный из проекта DeLM.
Продолжают ли результаты улучшаться при добавлении новых агентов?Да, но до определённого предела: средний процент прохождения тестов по пяти задачам ProgramBench вырос с 19,31% при одном агенте до 28,78% при 128 агентах (около 49% относительного улучшения), а большие команды также быстрее сходятся к результату.
Есть ли научная статья об этом проекте?Да, дизайн Agensh и эксперименты описаны в препринте, опубликованном на arXiv в сентябре 2026 года.
На каком бенчмарке его тестировали?Agensh тестировали на пяти самых сложных задачах ProgramBench, где агенты должны восстановить эталонные программы по наблюдениям за их поведением и прилагаемой документации, используя такие модели, как GPT-5.6-sol и GPT-6.1-sol.
Какие есть похожие проекты?DeLM, hive, ruflo, FlashRSI, openrig, rrsi, AI-Scientist-v3, context-language-models
Автор
microsoft
Размер
37 MB
Звёзды
926
Возраст
2 мес
Коммиты · 90 дней