Microsoft Research has released Agensh, an open-source framework designed to coordinate up to 1,024 LLM agents without a central orchestrator. The system allows agents to manage tasks through shared workspaces hosted on Gitea, communication via Mattermost, and a shared context board for tracking notes.
The framework operates by embedding the coordination loop directly into the prompt rather than runtime code. It is compatible with existing single-agent harnesses, such as the GitHub Copilot CLI, and supports integration with other tools through thin adapters.
Testing involved five ProgramBench tasks, including FFmpeg and PHP-src, with repositories containing up to 26,266 files. Using GPT-5.6-sol, the mean test-pass rate increased by 9.47 points, representing a 49% relative gain. For the pandoc task, performance scaled from 33.89% with one agent to 55.06% with 1,024 agents.
Speed metrics also showed improvement; for the pandoc task, 128 agents achieved a 30% pass rate within 30 minutes, while smaller teams of 32 and 8 agents required 60 and 90 minutes respectively. Newer tests with GPT-6.1-sol showed a 41.64% relative gain in pandoc scores when using 32 agents compared to one.
The release includes a single-node Docker Compose configuration and a baseline for single-agent setups. Technical requirements include Linux with Docker Engine, Compose v2, systemd, cgroup v2, and Python 3.11 or higher. The default configuration supports 8 workers and a 6-hour time limit per task.
