Microsoft Research Open-Sources Agensh, a Leaderless Harness That Ran 1,024 Coding Agents on One Program

Microsoft Research has published the code behind an experiment in which up to 1,024 AI coding agents rebuilt real software together with no agent in charge. The project, Agensh, went public on GitHub under the MIT licence, with a v1.0 release tagged on October 23 (UTC) and a README entry dated October 24, 2026, announcing the open-sourcing. The release includes the harness, experiment launch scripts and a reproduction guide. The accompanying preprint, “Agensh: Scaling Organizational Intelligence to 1,024 Agents”, was posted to arXiv on September 22. Its authors are Zhihao Zhan, Ting Song, Li Dong, Shaohan Huang, Jianxun Lian, Yan Xia and Furu Wei.

Most multi-agent coding systems use a planner or orchestrator that splits a job and hands out pieces. Agensh takes that role away. In the README’s words, “agents discover and claim sub-tasks, share findings, verify results and merge progress asynchronously, without a central orchestrator.”

Gitea, Mattermost and a shared board

The coordination runs on ordinary developer tools rather than a custom protocol. Each agent repeats the same loop: gather context, claim a sub-task, act, verify the result and merge. A self-hosted Gitea instance serves as the shared code workspace, Mattermost carries messages between agents, and a “board” service holds shared context and work claims. Microsoft says it adapted the board from DeLM, an open-source project. Inside each worker, the per-agent loop comes from GitHub’s Copilot CLI, driven through Agent Client Protocol (ACP) sessions.

The reproduction guide describes a team in which every agent has the same peer role. Agents merge through pull requests that need zero approvals, direct pushes to the main branch are blocked, and the CI merge gate is switched off for the experiments. A router service relays events and messages, and an idle watchdog nudges any worker that has been inactive for ten minutes. Each worker runs in a container limited to 64 GiB of memory, as an unprivileged user, with network access restricted to the model gateway.

Rebuilding Pandoc from the outside

Microsoft tested the harness on ProgramBench, a benchmark in which agents have to rebuild a reference program from its observable behaviour and bundled documentation. They can run the original binary but not read it. The team chose the five hardest tasks, which reimplement FFmpeg, GROMACS, Pandoc, PHP and Universal Ctags. Each run used GPT-5.6-sol at high reasoning effort with a six-hour budget and no internet access.

Microsoft’s numbers show gains from adding agents. Across the five tasks, the mean final test-pass rate rose from 19.31% with one agent to 28.78% with 128 agents, which the README describes as “approximately 49% relative”. Pandoc showed the largest effect. By Microsoft’s own figures, its pass rate went from 33.89% with a single agent to 50.94% with 128 and 55.06% with 1,024, all within the same six hours. That comparison is against one agent working alone, not against a coordinated team of the same size. The team says larger teams also reach comparable pass rates sooner. A newer model raised the ceiling. With GPT-6.1-sol, Pandoc’s pass rate went from 56.19% with one agent to 79.59% with 32, a 41.64% relative gain.

The researchers also report changes in how the agents behaved. As teams grew, “self-organized peer coordination, integration workflows and specialized roles gradually emerge,” the README says, even though every agent starts in the same peer role.

Not a weekend install

The public release runs on a single machine only. It ships a Docker Compose setup for a team of agents and a standalone single-agent baseline for comparison. The defaults are eight workers, six hours per task and a 272,000-token prompt window. The requirements are substantial: Linux with cgroup v2 and systemd, passwordless sudo for Docker, a separate ProgramBench checkout pinned to a specific commit, per-task container images, and a model gateway that the user supplies and pays for.

Microsoft says plainly that the release will not reproduce the paper’s figures exactly. It pins Copilot 1.0.70 and Mattermost MCP 0.5.1 and uses version 6 of ProgramBench’s cleanroom images. The original experiments used an older reference image whose binary differs “even when version strings match”. The guide asks anyone running it to “record its own revision rather than claim exact historical reproduction”. The 128- and 1,024-agent figures come from Microsoft’s original runs, while the public release covers only the single-node setup.

For teams building agent platforms, the design is the main thing to take away. Coordination runs through a Git forge and a chat server that people already use, and pass rates kept rising as Microsoft added agents. The figures, however, are Microsoft’s own and unreproduced, and they compare team sizes with a single agent, not with an orchestrated team of the same size.