What’s new here

Agensh has no lead process handing out work. All workers proceed asynchronously and share progress through shared infrastructure: a Gitea workspace, Mattermost messaging and a shared context mechanism. The authors report that coordination workflows and specialized roles emerge as the team grows.

What it does

Workers repeat one loop until the user’s goal is achieved: gather context, claim a sub-task, take action, verify results, merge progress. Workers run asynchronously. Gitea is the shared workspace, Mattermost is the messaging interface, and the shared context mechanism is adapted from DeLM. In the authors’ experiments, each agent’s own loop is the Copilot harness (github/copilot-cli).

The README reports results on the five hardest ProgramBench tasks, using GPT-5.6-sol at high reasoning effort, a six-hour budget and no Internet access. Agents rebuild reference programs from behavioral observations and bundled documentation. Mean test-pass rate goes from 19.31% with 1 agent to 28.78% with 128 agents. On Pandoc it goes from 33.89% with 1 agent to 50.94% with 128 and 55.06% with 1,024. With GPT-6.1-sol, Pandoc goes from 56.19% with 1 agent to 79.59% with 32 agents. These are the authors’ own numbers.

The release covers the harness, experiment launch scripts and a reproduction guide. It includes a single-node team setup using Docker Compose and a standalone single-agent baseline, both on native Copilot ACP sessions.

Who it’s for

Researchers and engineers who want to study how large agent teams coordinate without a manager, or who want to reproduce the paper’s scaling results. It is an experiment harness tied to a paper, so expect to read the reproduction guide before running anything.

Try it

The README has no short quick start. It points to REPRODUCE.md, which covers prerequisites, configuration, image builds, launch commands, evaluation and recovery. The team setup uses Docker Compose, and both setups use native Copilot ACP sessions. The entrypoints are scripts/launch_single_node.sh for the team setup and scripts/launch_single_agent.py for the single-agent baseline.

git clone https://github.com/microsoft/Agensh

Then follow REPRODUCE.md. There is a v1.0 release, but no prebuilt assets are listed.

How mature is it

Young. The repo was created on 2026-08-27, last pushed on 2026-10-23, and tagged v1.0 on 2026-10-23. It has 505 stars, 58 forks, 1 open issue, no open PRs and an MIT license. The GitHub API shows 1 contributor and 1 commit in the last 90 days, so this is a code release alongside a research paper (arXiv 2609.26781), not a long-maintained project. The README’s own scope note is a single-node setup.