Agensh

Agensh is a decentralized multi-agent harness open-sourced by Microsoft Research (MIT license). Its pitch: replace a central orchestrator with self-organization, and scale coding agents to a thousand concurrent workers.\n\nThe companion paper Agensh: Scaling Organizational Intelligence to 1,024 Agents (arXiv 2609.26781, 2026-09-22) is by Zhihao Zhan, Ting Song, Li Dong, Shaohan Huang, Jianxun Lian, Yan Xia, Furu Wei and coauthors. The GitHub repo microsoft/Agensh was created 2026-08-27; the v1.0 / open-source drop landed 2026-10-23–24 (README News dated Oct 24). It has about 293 stars and 35 forks. Project page: agens-harness.github.io/project/; short link aka.ms/Agensh.\n\nThe problem it targets: many multi-agent harnesses bottleneck on a central orchestrator. Agensh lets workers asynchronously claim sub-tasks and share progress through a light "organization infrastructure".

How it works\n\nAgensh defines a repeatable cooperation loop: workers concurrently gather context → claim a sub-task → act → verify → merge progress asynchronously until the user goal is met. The loop sits on three light infrastructure pieces:\n\n- Shared workspace — experiments use Gitea for proposed / in-progress / completed work\n- Message interface — Mattermost for worker-to-worker chat\n- Shared context — adapted from DeLM, keeping reusable findings and work intentions\n\nIn the paper and README evaluations, Agensh is typically stacked on each worker's own Copilot harness single-agent loop — the product-facing Copilot harness stays the per-agent loop while Agensh adds team scale.\n\n## Main results (author-reported)\n\nOn the five hardest ProgramBench tasks (FFmpeg, gromacs, pandoc, PHP-src, ctags) with GPT-5.6-sol (high), a six-hour budget and no Internet, mean final test-pass rises from 19.31% with 1 agent to 28.78% with 128 agents (~49% relative). On Pandoc: 33.89% → 50.94% → 55.06% at 1 → 128 → 1,024 agents. Larger teams also reach similar pass rates sooner. With GPT-6.1-sol (high), Pandoc six-hour pass rises from 56.19% (1 agent) to 79.59% (32 agents).\n\n> Figures are from the paper/README — not independently reproduced here; for reference only.

Get it and reproduce\n\n- Code: https://github.com/microsoft/Agensh (MIT)\n- Project page & demo: https://agens-harness.github.io/project/\n- Paper: https://arxiv.org/abs/2609.26781\n- Short link: https://aka.ms/Agensh\n\nThe release includes a Docker Compose single-node team setup, a standalone single-agent baseline, and a reproduction guide (prereqs, config, image builds, launch commands, evaluation, recovery). Experiment entrypoints are the team launcher and single-agent launcher in the repo.\n\n## Who it's for\n\n- Teams studying multi-agent scalability without a central orchestrator\n- Research and engineering experiments that need many coding agents under a hard latency / time budget\n\n> This is a research harness, not a claim that Microsoft shipped a new consumer coding product.