{"slug": "show-hn-agentsh-a-decentralized-scalable-multiagent-harness", "title": "Show HN: Agentsh – A Decentralized, Scalable Multiagent Harness", "summary": "Microsoft open-sourced Agensh, a decentralized multi-agent harness that scales to 1,024 agents without a central orchestrator, on 2026-10-24 alongside experiment launch scripts and a reproduction guide. On the five hardest ProgramBench tasks using GPT-5.6-sol at high reasoning effort with a six-hour budget and no Internet access, the mean final test-pass rate rose from 19.31% with 1 agent to 28.78% with 128 agents, a roughly 49% relative improvement; on Pandoc the rate rose from 33.89% with 1 agent to 50.94% with 128 agents and 55.06% with 1,024 agents. With GPT-6.1-sol at high reasoning effort, the six-hour Pandoc test-pass rate rose from 56.19% with 1 agent to 79.59% with 32 agents, a 41.64% relative improvement.", "body_md": "🚀 **2026-10-24:** We open-sourced **Agensh**, including the harness, experiment launch\nscripts, and [reproduction guide](https://github.com/microsoft/Agensh/blob/main/REPRODUCE.md).\n\n📢 **2026-09-22:** We published the **Agensh** [preprint](https://arxiv.org/abs/2609.26781) on arXiv.\n\n**Agensh is a self-organized multi-agent harness that scales to 1,024 agents.**\nAgents discover and claim sub-tasks, share findings, verify results and\nmerge progress asynchronously, without a central orchestrator. A shared\nworkspace, message interface and shared context connect each agent's local\nprogress to the team.\n\nWe evaluated Agensh on the five hardest ProgramBench tasks\nusing GPT-5.6-sol with high reasoning effort, a six-hour budget and no Internet access.\nAgents rebuild reference programs from behavioral observations and bundled\ndocumentation. Across the 5 evaluated tasks, the mean final test-pass rate rises\nfrom **19.31% with 1 agent** to **28.78% with 128 agents**, an improvement of\napproximately **49% relative**. Larger teams also reach comparable pass rates\nsooner. On **Pandoc**, the final test-pass rate rises from **33.89% with 1 agent**\nto **50.94% with 128 agents** and **55.06% with 1,024 agents** under the same\nsix-hour budget. We find that self-organized peer coordination, integration\nworkflows and specialized roles gradually emerge as the team grows.\n\nWith the latest GPT-6.1-sol at high reasoning effort, the six-hour Pandoc test-pass\nrate rises from **56.19% with 1 agent** to **79.59% with 32 agents**, a relative\nimprovement of **41.64%**.\n\n**Agensh defines the multi-agent cooperation loop.**\nWorkers concurrently gather context, claim sub-tasks, take action, verify results\nand merge progress in a self-organized manner, repeating until the user’s goal\nis achieved. All workers proceed asynchronously and share progress through the\ninfrastructure.\n[Gitea](https://about.gitea.com/) provides the shared workspace,\n[Mattermost](https://mattermost.com/) provides the messaging interface, and the\nshared context mechanism is adapted from [DeLM](https://github.com/yuzhenmao/DeLM).\nIn our experiments, we integrate Agensh on top of the per-agent agentic loop\nof [Copilot harness](https://github.com/github/copilot-cli).\n\nThis release includes a single-node team setup using Docker Compose and a standalone single-agent baseline, both using native Copilot ACP sessions.\n\n- [Reproduction guide](https://github.com/microsoft/Agensh/blob/main/REPRODUCE.md) : prerequisites, configuration, image builds,\nlaunch commands, evaluation, recovery and release differences.\n- [Repository layout](https://github.com/microsoft/Agensh/blob/main/REPRODUCE.md#repository-layout) : component and folder guides.\n- [Team launcher](https://github.com/microsoft/Agensh/blob/main/scripts/launch_single_node.sh) and[single-agent launcher](https://github.com/microsoft/Agensh/blob/main/scripts/launch_single_agent.py) : experiment entrypoints.\n\n```\n@article{zhan2026agensh,\n  title   = {Agensh: Scaling Organizational Intelligence to 1,024 Agents},\n  author  = {Zhan, Zhihao and Song, Ting and Dong, Li and Huang, Shaohan and Lian, Jianxun and Xia, Yan and Wei, Furu},\n  journal = {arXiv preprint arXiv:2609.26781},\n  year    = {2026},\n  url     = {https://arxiv.org/abs/2609.26781}\n}\n```\n\nCode is licensed under the [MIT License](https://github.com/microsoft/Agensh/blob/main/LICENSE). See the\n[contribution guidelines](https://github.com/microsoft/Agensh/blob/main/REPRODUCE.md#contributing),\n[security policy](https://github.com/microsoft/Agensh/blob/main/SECURITY.md),\n[support information](https://github.com/microsoft/Agensh/blob/main/REPRODUCE.md#security-and-support), and\n[trademark guidance](https://github.com/microsoft/Agensh/blob/main/REPRODUCE.md#trademarks).", "url": "https://wpnews.pro/news/show-hn-agentsh-a-decentralized-scalable-multiagent-harness", "canonical_source": "https://github.com/microsoft/Agensh", "published_at": "2026-10-24 18:19:35+00:00", "updated_at": "2026-10-24 18:45:12.753315+00:00", "lang": "en", "topics": ["ai-agents", "artificial-intelligence", "ai-research", "large-language-models", "ai-tools"], "entities": ["Microsoft", "Agensh", "GPT-5.6-sol", "GPT-6.1-sol", "ProgramBench", "Pandoc", "Gitea", "Mattermost"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/show-hn-agentsh-a-decentralized-scalable-multiagent-harness", "markdown": "https://wpnews.pro/news/show-hn-agentsh-a-decentralized-scalable-multiagent-harness.md", "text": "https://wpnews.pro/news/show-hn-agentsh-a-decentralized-scalable-multiagent-harness.txt", "jsonld": "https://wpnews.pro/news/show-hn-agentsh-a-decentralized-scalable-multiagent-harness.jsonld"}}