# Agensh (Microsoft Research multi-agent harness for coding agents)

> Updated 2026-10-25 · type: tool · category: mcp-agents · status: active · rev 1

Agensh is Microsoft Research's MIT harness where many coding agents coordinate as peers over a shared repo, chat bus and board, with no lead agent.

- Open source: yes (MIT)
- Self-hostable: yes
- Pricing: Free (MIT), no hosted tier; real cost is LLM inference for many concurrent agents plus the Docker host.
- Best for: Engineering and GTM-engineering teams experimenting with many parallel coding agents on one large build, who can run Linux infrastructure and pay for heavy LLM inference.
- Not for: Teams wanting a turnkey agent product: it needs Linux, systemd, cgroup v2, passwordless sudo and a Responses-API upstream with a 272k context window, and the public release is single-node only.
- Last verified: 2026-10-25

- **Canonical:** https://gtmstacker.com/registry/tool/agensh/
- **Source:** [Microsoft just open-sourced Agensh, a multi-agent harness that runs up to 1,024 coding agents with no lead agent in charge](https://x.com/Marktechpost/status/2114200105108586623) — Marktechpost
- **Tags:** mcp-agents, ai-infrastructure, coding-agents, multi-agent, agent-orchestration, microsoft, research, self-hosted, open-source
- **Repository:** https://github.com/microsoft/Agensh

## Is Agensh (Microsoft Research multi-agent harness for coding agents) open source?

Yes, Agensh (Microsoft Research multi-agent harness for coding agents) is open source under the MIT license.

## How much does Agensh (Microsoft Research multi-agent harness for coding agents) cost?

Free (MIT), no hosted tier; real cost is LLM inference for many concurrent agents plus the Docker host. (free).

## Can I self-host Agensh (Microsoft Research multi-agent harness for coding agents)?

Yes, Agensh (Microsoft Research multi-agent harness for coding agents) can be self-hosted (the source is available under the MIT license).


---

**Agensh** (`microsoft/Agensh`) is an **MIT**-licensed multi-agent harness from **Microsoft Research**. Instead of one lead agent handing out tasks, a team of coding agents works as peers: they share a **Gitea** workspace, talk over a **Mattermost** message bus, and keep shared context on a board service, with each agent running as a GitHub Copilot ACP session.

## Provenance

- SPDX **MIT**, confirmed by the GitHub license API (`license.spdx_id`) and the README ("Code is licensed under the MIT License").
- Identity: official `microsoft` org, not a fork, not archived. The README cross-links the paper (arXiv 2609.26781, "Agensh: Scaling Organizational Intelligence to 1,024 Agents") and the project page aka.ms/Agensh. No sign of a lookalike repo.
- Activity as of 2026-10-25 (GitHub API): **192 stars, 21 forks**. Repo created 2026-08-27; first commit "Initial release of Agensh" on 2026-10-22; last push 2026-10-23. README news line: open-sourced 2026-10-24.
- Self-hosting: yes, but heavy. The public release is a single-node Docker Compose setup plus a standalone single-agent baseline. It needs Linux, Docker Engine with Compose v2, systemd, cgroup v2, passwordless sudo, and your own `LLM_UPSTREAM_TOKEN` for a Responses-API model with a 272k-token window.
- Pricing: free software, no hosted tier. The cost is model inference across the concurrent agents plus the host.
- **Vendor claim:** on the five hardest ProgramBench tasks (GPT-5.6-sol, six hours, no internet), mean test-pass rate rose from 19.31% with one agent to 28.78% with 128 agents. On Pandoc it went from 33.89% (1 agent) to 55.06% (1,024 agents). These are Microsoft Research's own numbers, not independently reproduced, and the 1,024-agent run is not what the public release ships.
- Surfaced via an X post in the 2026-10-25 pull (the post itself returned HTTP 402 and was not used as evidence). Curated and verified in run `2026-10-25-registry-update-01` (step-3 verification, 2026-10-25).

## Why it matters for a GTM stack

Most agent setups put one orchestrator in charge and the rest wait for instructions. Agensh tests the other shape: a flat team that coordinates through the same tools a human team uses, a repo, a chat channel and a shared board. For GTM engineering teams building internal tools with agents, it is a readable, MIT-licensed reference for that pattern. The honest read: this is a research harness, not a product. Expect real Linux setup work and a large inference bill before it pays off.