Microsoft Research has published Agensh, spelled that way on GitHub (`microsoft/Agensh`), as MIT-licensed open source: a self-organized multi-agent harness that scales workers without a central orchestrator. The repository README’s News line dates the open-source drop 2026-10-24; GitHub’s `v1.0` release shows `published_at` 2026-10-23T08:03:16Z (about 16:03 Beijing time on October 23, 2026). The accompanying paper, *Agensh: Scaling Organizational Intelligence to 1,024 Agents*, is on arXiv as 2609.26781 with a 2026-09-22 preprint date, and Microsoft Research lists a matching publication entry. This note is not a same-day first-release story.

What Agensh is claiming

Agensh runs concurrent workers that gather context, claim sub-tasks, act, verify, and merge progress asynchronously through a shared workspace (Gitea), messaging (Mattermost), and shared context. Experiments in the README stack that loop on top of each worker’s Copilot harness single-agent session—so the product-facing Copilot harness stays the per-agent loop while Agensh adds team scale. That is a research harness, not a claim that Microsoft shipped a new consumer coding product.

On the five hardest ProgramBench tasks, with GPT-5.6-sol at high reasoning effort, a six-hour budget, and no Internet, the authors report mean final test-pass rising from 19.31% with 1 agent to 28.78% with 128 agents (about a 49% relative lift). On Pandoc, they report 33.89% → 50.94% at 128 agents → 55.06% at 1,024 agents under the same six-hour budget. With GPT-6.1-sol (high), the README adds a Pandoc six-hour figure from 56.19% with 1 agent to 79.59% with 32 agents (about 41.64% relative). Larger teams also reach similar pass rates sooner, per the authors. We have not reproduced these runs; for reference only.

How this sits next to “what a harness buys”

A companion angle—same model, different coding harness, and how the bill moves when pass rate barely does—is covered in Same Model, Different Coding Harness: Pass Rate Barely Moves, the Bill Does. Agensh’s pitch is the other axis: keep the per-agent Copilot loop, change how many peers share the job. Install-oriented DeepSeek Harness steps are in our DeepSeek Harness Install Guide: Windows, Linux, CLI, Plugins; VS Code’s product-side Copilot harness default is in VS Code 1.143 Makes the Copilot Harness the Default for New Sessions. For another multi-agent scheduling caveat on OpenAI’s Agents API, see Why Agents API Scheduled Automations Show Completed Before Work Finishes.

Whether Agensh’s ProgramBench curves hold on a given repo, model, and time budget is for reference only. Feedback can go through the site-owner contact link in the footer.