🚀 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.
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 the latest 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%.
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 experiments, we integrate Agensh on top of the per-agent agentic loop of Copilot harness.
This release includes a single-node team setup using Docker Compose and a standalone single-agent baseline, both using native Copilot ACP sessions.
- Reproduction guide : prerequisites, configuration, image builds, launch commands, evaluation, recovery and release differences.
- Repository layout : component and folder guides.
- Team launcher andsingle-agent launcher : experiment entrypoints.
@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}
}
Code is licensed under the MIT License. See the contribution guidelines, security policy, support information, and trademark guidance.