# Microsoft open-sources Agensh, a multi-agent harness its authors scaled to 1,024 agents

Canonical: https://theagenticweekly.com/news/microsoft-agensh-multi-agent-harness-1024-agents · Evidence: Reported · First seen: 2026-10-26 03:47 UTC · Source published: 2026-10-23 08:03 UTC · Brief published: 2026-10-26 03:51 UTC · Updated: 2026-10-26 03:51 UTC · Author: Agentic Weekly

> Microsoft has open-sourced Agensh, a multi-agent harness with no central orchestrator that its authors report running with up to 1,024 agents; the benchmark numbers are the authors' own and we have not reproduced them.

## What changed

- Microsoft published Agensh, described as "a decentralized, scalable multiagent harness", under the MIT license. GitHub shows a v1.0 release on 2026-10-23 (UTC); the README dates the open-source announcement 2026-10-24.
- Design: agents discover and claim sub-tasks, share findings, verify results and merge progress asynchronously, with no central orchestrator. Gitea provides the shared workspace and Mattermost the messaging interface.
- In the authors' experiments Agensh sits on top of the per-agent loop of the Copilot harness (github/copilot-cli), and its shared-context mechanism is adapted from DeLM.
- Authors' results on ProgramBench: across five tasks the mean final test-pass rate rose from 19.31% with 1 agent to 28.78% with 128 agents; on Pandoc it rose from 33.89% with 1 agent to 55.06% with 1,024 agents under a six-hour budget.
- The release includes a single-node team setup using Docker Compose and a single-agent baseline, with a reproduction guide in the repository.

## In the source's words

> "Agensh is a self-organized multi-agent harness that scales to 1,024 agents."
> (Agensh README, https://github.com/microsoft/Agensh)

> "Agents discover and claim sub-tasks, share findings, verify results and merge progress asynchronously, without a central orchestrator."
> (Agensh README, https://github.com/microsoft/Agensh)

## What we have not verified

- We did not run Agensh or reproduce any benchmark number; all results are the authors' own, run with models and settings described in their README.
- We found it through a press feed first; the vendor's own repository is the primary source for every fact here.
- The README's announcement date (2026-10-24) and GitHub's release date (2026-10-23 UTC) differ; we show the GitHub date as the source publish time.

## Timeline

- Source published: 2026-10-23 08:03 UTC
- First seen by our watcher: 2026-10-26 03:47 UTC (via MarkTechPost feed (watcher seed run); primary source is the Microsoft GitHub repository), 2.8 days after the source's timestamp
- Brief published: 2026-10-26 03:51 UTC

## Sources

- [microsoft/Agensh README (GitHub)](https://github.com/microsoft/Agensh) (primary source) (fetched 2026-10-26)
- [microsoft/Agensh v1.0 release (GitHub)](https://github.com/microsoft/Agensh/releases/tag/v1.0) (fetched 2026-10-26)
- [MarkTechPost coverage (how we first saw it)](https://www.marktechpost.com/2026/10/24/microsoft-agensh-multi-agent-harness-1024-agents/) (fetched 2026-10-26)

## Related

- [All news briefs](https://theagenticweekly.com/news)
- [This week's releases](https://theagenticweekly.com/this-week)
- [New AI models this week](https://theagenticweekly.com/models/this-week)

## How to cite

Agentic Weekly, "Microsoft open-sources Agensh, a multi-agent harness its authors scaled to 1,024 agents", published 2026-10-26 03:51 UTC, https://theagenticweekly.com/news/microsoft-agensh-multi-agent-harness-1024-agents