Microsoft Open-Sources Agensh: A Multi-Agent Harness for 1,024 Agents

Why it matters

Microsoft Research has released Agensh, an open-source harness that coordinates very large teams of coding agents without any central orchestrator. Instead of a lead agent that plans, assigns, and merges work, each worker runs the same prompt-driven loop: it gathers context, claims a sub-task, acts, verifies against acceptance criteria, and merges into a shared Git repository. Coordination happens through three services: Gitea for the shared workspace, Mattermost for messaging, and a shared findings board holding typed notes. The work targets what the researchers call the orchestrator ceiling, a limitation they identify in frameworks like Claude Code agent teams, Codex sub-agents, Copilot fleet, and Kimi Agent Swarm. On ProgramBench's hardest tasks, adding agents improved both final pass rates and the speed at which targets were reached, with the strongest results on pandoc using GPT-6.1-sol. Notably, roles such as reviewers, trusted integrators, and recovery helpers emerged spontaneously as team size grew, rather than being hard-coded. The code is MIT licensed, though token costs remain undisclosed.