Microsoft unveils Agensh: a new framework for managing 1,024 AI agents

Microsoft Research has released Agensh, a decentralized framework designed to manage large teams of AI agents. The system eliminates the need for a central orchestrator, allowing each agent to independently select tasks, share findings, and merge code into a shared Git repository. Agens, which operates on a 5-step loop, enables agents to coordinate without a lead agent, drawing inspiration from existing frameworks like Claude Code and Codex. The framework utilizes three services—Gitea, Mattermost, and a shared context board—to facilitate communication and collaboration. Testing on ProgramBench showed significant improvements in performance, with a 9.47-point increase in test-pass rates when scaling from 1 to 1,024 agents. The latest results using GPT-6.1-sol showed a 41.64% relative gain in performance on the pandoc task. The release includes Docker Compose setups and a reproduction guide for users to experiment with the framework. The framework's ability to scale efficiently and its integration with GitHub Copilot CLI and other models like Claude Code highlight its potential for large-scale AI collaboration. Microsoft has also provided resources for developers, including a single-node Docker Compose team and a single-agent baseline. The release marks a significant step in decentralized AI coordination, offering a new approach to managing complex multi-agent systems.