Agensh: Build a Leaderless Agent Team in Python

On October 24, Microsoft Research open-sourced Agensh under the MIT license: a harness that runs up to 1,024 coding agents against one repository with no lead agent and no scheduler handing out work. It was one of the most-shared agent stories of the weekend, and the reason is the headline result. On the Pandoc rebuild task, pass rate went from 33.89% with one agent to 50.94% with 128 and 55.06% with 1,024. Then came the twist that got people arguing: in a second run, a single agent on the newer GPT-6.1-sol scored 56.19%, higher than the entire 1,024-agent fleet on the older model.

This guide does two things. First it explains exactly how Agensh coordinates agents and how to launch it from the official reproduction guide. Then it has you build a small, runnable version of its core idea in Python (a shared board, atomic task claims, typed FAIL notes, a verify gate and a serialized merge) so you can see the pattern work on your laptop in under a minute, with no Docker and no API key.