Microsoft Analysis has open-sourced Agensh, a decentralized harness for giant groups of coding brokers. The Agensh multi-agent harness removes the central orchestrator that the majority agent frameworks depend upon. Each employee picks its personal sub-task, shares findings and merges code into a typical Git repo. On ProgramBench, the crew scaled the setup from 1 agent to 1,024 brokers.
TL;DR
- A Harness: Examined from 1 to 1,024 brokers. Every employee will get a 272,000-token enter and 128,000-token output restrict.
- Runs on: Linux with Docker Compose. Single node by default (8 staff). The 1,024-agent run used 16 nodes, 64 brokers every. Staff have a 64 GiB reminiscence ceiling.
- Efficiency: extra brokers raised scores on all 5 hardest ProgramBench duties, and reached them sooner.
- Greatest: 79.59% pandoc cross charge with 32 brokers on GPT-6.1-sol. GPT-5.6-sol hit 55.06% with 1,024 brokers.
- Worst: the 5-task imply reached solely 28.78% at 128 brokers. Going from 1 to eight brokers added 1.37 factors.
What’s Agensh?
Agensh is a multi-agent harness from Microsoft Analysis that coordinates many LLM brokers with no lead agent.
The goal drawback is the orchestrator ceiling. In orchestrator-worker designs, a lead agent plans, assigns and merges all work. The research paper cites Claude Code agent groups, Codex sub-agents, Copilot fleet and Kimi Agent Swarm as examples. Agensh lets the employees do this coordination themselves.
How does the Agensh cooperation loop work?
Every employee runs the identical 5-step loop, concurrently and asynchronously:
- Collect context: learn the objective, peer progress, messages and findings.
- Declare a sub-task: put up a CLAIM, then settle overlaps by direct message.
- Take motion: work domestically with instruments and put up helpful findings early.
- Confirm outcomes: verify the work in opposition to the sub-task’s acceptance standards.
- Merge progress: merge into the shared repo, put up an replace, then repeat.
Three companies join the employees. Gitea hosts the shared workspace, with branches, PRs and merge conflicts. Mattermost carries a crew channel plus direct messages for pressing collisions. A shared context board, tailored from DeLM, shops typed notes: OBSERVED, FACT, FAIL, CLAIM and PATCH_SUMMARY.
The loop lives within the immediate, not in runtime code. Each employee will get the identical immediate besides its ID. Agensh sits on prime of a single-agent harness. The experiments used GitHub Copilot CLI. The analysis paper states skinny adapters can join different harnesses, comparable to Claude Code.
(operate(){var f=doc.getElementById(“agx-frame”);window.addEventListener(“message”,operate(e){if(e.information&&e.information.agxHeight&&f&&e.supply===f.contentWindow){f.model.top=e.information.agxHeight+”px”;}});})();
How nicely does Agensh scale on ProgramBench?
Microsoft crew examined the 5 hardest ProgramBench duties: FFmpeg, gromacs, pandoc, PHP-src and ctags. Brokers rebuild every program from its noticed habits. They work offline, inside 6 hours. Reference repos vary from 2,767 recordsdata (pandoc) to 26,266 recordsdata (PHP-src).
With GPT-5.6-sol at excessive reasoning effort, the imply last test-pass charge moved as follows:
- 1 agent: 19.31%
- 8 brokers: 20.68%
- 32 brokers: 26.52%
- 128 brokers: 28.78%
That may be a 9.47-point achieve, about 49% relative. On pandoc alone, scores had been 33.89% (1), 41.75% (8), 43.11% (32), 50.94% (128) and 55.06% (1,024).
Velocity improved too: on pandoc, 128 brokers handed 30% on the 30-minute checkpoint. 32 and eight brokers wanted 60 and 90 minutes.
The README provides a more recent mannequin. With GPT-6.1-sol, pandoc rose from 56.19% (1 agent) to 79.59% (32 brokers). That may be a 41.64% relative achieve. These GPT-6.1-sol figures should not within the v1 paper.
What behaviors emerged because the crew grew?
The analysis crew traced employee trajectories at every scale:
- 8 brokers: friends agreed on interfaces and resolved overlapping claims themselves.
- 32 brokers: friends reviewed shared work; one approval was withdrawn after a counterexample.
- 128 brokers: staff reused trusted reviewers and standardized an integration protocol.
- 1,024 brokers: a number of staff acted as integrators, and friends recovered failed work.
None of those roles was hard-coded.
How does Agensh evaluate with different multi-agent harnesses?
| Characteristic | Agensh | Claude Code agent teams | Kimi Agent Swarm | Copilot CLI /fleet |
|---|---|---|---|---|
| Coordination | No orchestrator; friends self-claim work | Mounted lead session coordinates teammates | Orchestrator decomposes tasks | Copilot assigns components to subagents |
| Brokers proven | 1,024 examined | No onerous restrict; 3 to five advisable | As much as 100 sub-agents | Not disclosed |
| Shared state | Gitea repo + shared context board | Shared job checklist with file locking | Not disclosed | Plan from plan mode |
| Messaging | Mattermost channel + direct messages | Mailbox messages between teammates | Not disclosed | Not disclosed |
| Fashions | Mannequin agnostic; examined GPT-5.6-sol, GPT-6.1-sol | Claude fashions | Kimi K2.5 | Not disclosed |
| Velocity declare | 30% on pandoc at 30 min (128 brokers) | Not disclosed | 4.5x sooner than sequential (self-reported) | Not disclosed |
| Availability | Open supply, MIT | In Claude Code; experimental, off by default | High-tier subscribers, analysis preview | In Copilot CLI |
How do you run Agensh?
The discharge features a single-node Docker Compose crew and a single-agent baseline. Each use native Copilot ACP periods. You want Linux with Docker Engine, Compose v2, systemd, cgroup v2 and Python 3.11+. Defaults are 8 staff, 6 hours per job and gpt-5.6-sol at excessive effort. You additionally want ProgramBench (v1.2.5 was validated) and a mannequin API token. Full steps sit within the reproduction guide.
Key Takeaways
- Agensh coordinates as much as 1,024 brokers with no central orchestrator.
- Staff self-claim duties through Gitea, Mattermost and a shared findings board.
- 128 brokers lifted the 5-task ProgramBench imply from 19.31% to twenty-eight.78%.
- GPT-6.1-sol with 32 brokers hit 79.59% on pandoc.
- Code is MIT licensed, however token value isn’t disclosed.
Try the Paper, GitHub Repo and Project Page. Additionally, be at liberty to observe us on Twitter and don’t overlook to hitch our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.
Must associate with us for selling your GitHub Repo OR Hugging Face Web page OR Product Launch OR Webinar and many others.? Connect with us
The put up Microsoft Open-Sources Agensh: A Multi-Agent Harness for 1,024 Agents appeared first on MarkTechPost.

