Microsoft Research has open-sourced Agensh, a multi-agent harness that coordinates up to 1,024 coding agents without a central orchestrator, in a release dated October 24, 2026. According to the project’s README, the release includes the code, experiment launch scripts and reproduction guide, under the MIT licence. Workers claim their own sub-tasks, share findings and merge progress through a shared workspace, a messaging channel and a shared context board. In tests on the five hardest ProgramBench tasks, scaling from one agent to 128 lifted the mean test-pass rate from 19.31% to 28.78%. On pandoc alone, 1,024 agents rebuilding the program from scratch reached 55.06%, up from 33.89% for a single agent.
How Agensh Works Without a Central Orchestrator
Most multi-agent systems give one lead agent the job of planning, assigning work and merging results. Agensh removes that role. Every worker runs the same five-step loop: gather context, claim a sub-task, take action, verify the result and merge progress. Workers run concurrently and asynchronously, and they repeat the loop until the goal is met.
Three shared services connect the workers. Gitea hosts the shared workspace, with branches, pull requests and merge conflicts. Mattermost carries a team channel plus direct messages for urgent collisions. A shared context board, adapted from DeLM, stores typed notes: OBSERVED, FACT, FAIL, CLAIM and PATCH_SUMMARY. The loop itself lives in the prompt, not in runtime code: every worker gets the same prompt except for its own ID.
Agensh sits on top of a single-agent harness. The experiments used GitHub Copilot‘s command-line harness through native Copilot ACP sessions. The paper says thin adapters can connect other harnesses, such as Claude Code. That design keeps the cooperation logic portable: swapping the underlying agent tool does not change how workers coordinate.
What 1,024 Agents Achieved on ProgramBench
The team tested Agensh on the five hardest ProgramBench tasks: FFmpeg, gromacs, pandoc, PHP-src and ctags. Agents rebuild each program from observed behaviour and bundled documentation. They worked offline, within a six-hour budget. Reference repositories ranged from 2,767 files for pandoc to 26,266 files for PHP-src.
With GPT-5.6-sol at high reasoning effort, the mean final test-pass rate moved from 19.31% with one agent to 20.68% with 8, 26.52% with 32 and 28.78% with 128. On pandoc alone, scores rose from 33.89% with one agent to 41.75% with 8, 43.11% with 32, 50.94% with 128 and 55.06% with 1,024. Larger teams also reached a given score sooner: on pandoc, 128 agents passed 30% at the 30-minute checkpoint, while 32 agents needed 60 minutes and 8 agents needed 90.
A newer model pushed pandoc further. With GPT-6.1-sol, the pandoc score rose from 56.19% with one agent to 79.59% with 32 agents, a relative gain of 41.64%. MarkTechPost notes these GPT-6.1-sol figures come from the README, not from the v1 paper. The 1,024-agent run used 16 nodes with 64 agents each.
Behaviour changed as teams grew. With 8 agents, peers agreed on interfaces and resolved overlapping claims themselves. With 32, peers reviewed shared work. With 128, workers reused trusted reviewers and standardised an integration protocol. With 1,024, several workers acted as integrators, and peers recovered failed work. None of these roles was hard-coded.
Running Agensh: Requirements and Limits
The release includes a single-node Docker Compose team and a single-agent baseline. By default the single node runs 8 workers, with six hours per task and GPT-5.6-sol at high effort. The setup needs Linux with Docker Engine, Compose v2, systemd, cgroup v2 and Python 3.11 or later. ProgramBench version 1.2.5 was validated, and a model API token is required.
Each worker gets a 272,000-token input limit and a 128,000-token output limit, with a 64 GiB memory ceiling. Token costs are not disclosed, so a team reproducing the 1,024-agent run cannot yet budget it from the release alone. The code is MIT-licensed, so teams can use, modify and redistribute it freely.
Why It Matters
Agensh matters because it tests a different scaling lever. Instead of waiting for a stronger model, teams can ask whether more agents, organised without a leader, do better work on hard, long-horizon tasks. The pandoc results suggest the answer is yes up to a point: gains flattened between 128 agents on the five-task mean, yet kept climbing on pandoc itself. That split is the honest signal in the release: agent count helps most where work splits cleanly and verification is cheap.
The release also lands beside rival approaches. Google’s Harness2 framework improves the harness around a frozen model, while the AWS Infinity runtime parks idle agents to run tens of thousands at once. Agensh attacks the coordination problem instead: how workers divide labour without a boss. For teams building coding fleets on Copilot or Claude Code, it offers a reproducible baseline, and a warning that token bills, not coordination, may set the practical ceiling.
Background
Agensh comes from Microsoft Research authors Zhihao Zhan, Ting Song, Li Dong, Shaohan Huang, Jianxun Lian, Yan Xia and Furu Wei, who published the preprint on September 22, 2026, as arXiv paper 2609.26781. The work joins a year in which agent builders have repeatedly found that identical models perform very differently depending on the scaffolding around them, and open-sourcing the harness lets other teams test the team’s scaling claims rather than take them on trust.
Source: Microsoft Research, Agensh repository on GitHub; MarkTechPost, October 25, 2026
0 Comments