Microsoft open-sources Agensh, a 1,024-agent coordination harness
A decentralised multi-agent framework with no central orchestrator, letting coding agents claim sub-tasks and merge work themselves at a scale most harnesses do not attempt.
Microsoft open-sourced Agensh on 24 October, a multi-agent harness built to coordinate up to 1,024 AI agents on one task without a central orchestrator. Agents discover and claim sub-tasks themselves, using Gitea as a shared workspace and Mattermost for messaging, then verify and merge each other’s progress asynchronously rather than waiting on a coordinator to assign work.
Microsoft’s own benchmarks, run on ProgramBench tasks with GPT-5.6-sol, show the mean test-pass rate rising from 19.31% with a single agent to 28.78% with 128 agents. On the Pandoc task specifically, it climbs further, to 55.06% with the full 1,024-agent team. The project is MIT licensed, and the underlying research was posted to arXiv in September.
Why it matters: most agent frameworks still assume one agent per task, or a central coordinator that becomes the bottleneck once you add more agents. Agensh is a working, tested alternative pattern for running much larger agent swarms on hard problems, and it’s open for anyone building agent infrastructure to read or adopt directly rather than re-derive.
The caveat: the gains are not uniform across models or tasks. With GPT-6.1-sol on the same Pandoc task, the jump from 1 to 32 agents already reaches 79.59%, suggesting returns flatten well before 1,024 agents for stronger models.