Agensh is an open-source multi-agent harness from Microsoft researchers in which a team of coding agents organizes itself without a central orchestrator.
Follow Agensh to get an email when its listing changes, such as new pricing or features. One digest a day at most, and you can unsubscribe anytime.
Each agent gathers context, claims a sub-task, acts, verifies the result and merges progress asynchronously, with a Gitea instance as the shared workspace, Mattermost for messaging and a shared context service adapted from DeLM. The released setup runs every worker on GitHub Copilot's agent loop through native ACP sessions.
- 01
Self-organized cooperation loop: agents discover and claim sub-tasks, share findings, verify results and merge progress without a central orchestrator
- 02
Shared Gitea workspace, Mattermost messaging and a shared context and work-claim board service
- 03
Workers built on GitHub Copilot's agentic loop through native ACP sessions
- 04
Single-node team launcher using Docker Compose plus a standalone single-agent baseline
- 05
Reproduction guide, experiment launch scripts and ProgramBench scoring tools
- 06
Scaled to 1,024 concurrent agents in the authors' experiments
Open Source
Free and open source under the MIT license; you supply GitHub Copilot access for the worker sessions plus a Linux host with Docker to run the team.
Best for
Researchers and platform engineers studying how large teams of coding agents coordinate, and anyone who wants to reproduce or extend the Agensh ProgramBench experiments on their own hardware.
Not ideal for
Developers looking for an everyday coding assistant; the release is a research harness that needs Linux with Docker Engine, Compose v2, systemd, cgroup v2 and passwordless sudo, and its defaults run long, token-heavy six-hour tasks.
A rare open, reproducible look at agent teams at the scale of hundreds of workers, worth studying for anyone designing multi-agent coding systems, but a research release rather than a product you install for daily work.
Open-sourced in October 2026 (release v1.0) with the harness, launch scripts and a reproduction guide, following the September 2026 arXiv preprint. The authors report that on the five hardest ProgramBench tasks the mean test-pass rate rose from 19.31% with 1 agent to 28.78% with 128 agents; these are the authors' own results.
As of
Is Agensh free?+
Yes - Agensh is Open Source. Free and open source under the MIT license; you supply GitHub Copilot access for the worker sessions plus a Linux host with Docker to run the team.
Is Agensh open source?+
Yes - Agensh is open source. Free and open source under the MIT license; you supply GitHub Copilot access for the worker sessions plus a Linux host with Docker to run the team.
Who is Agensh best for?+
Researchers and platform engineers studying how large teams of coding agents coordinate, and anyone who wants to reproduce or extend the Agensh ProgramBench experiments on their own hardware.
Who is Agensh not ideal for?+
Developers looking for an everyday coding assistant; the release is a research harness that needs Linux with Docker Engine, Compose v2, systemd, cgroup v2 and passwordless sudo, and its defaults run long, token-heavy six-hour tasks.
What are the best Agensh alternatives?+
The closest Agensh alternatives on ai.dosa.dev are Jules, OpenHands, SWE-agent - all listed under Autonomous & Async Agents.
Who makes Agensh?+
Agensh is developed by Microsoft. It is listed in the Autonomous & Async Agents category on ai.dosa.dev.
Favorite this tool to revisit it later, or Zap it to contribute to the public vote count.