Microsoft Research Open-Sources Agensh, a Decentralized Harness for 1,024 Coding Agents

Agensh targets what the research paper calls the orchestrator ceiling. In orchestrator-worker designs, a lead agent plans, assigns and merges all work; the paper cites Claude Code agent teams, Codex sub-agents, Copilot fleet and Kimi Agent Swarm as examples. Agensh instead has workers coordinate themselves through a five-step asynchronous loop: gather context from the goal, peer progress, messages and findings; post a CLAIM for a sub-task and settle overlaps by direct message; work locally with tools and post useful findings early; verify results against acceptance criteria; merge into the shared repository, post an update and repeat. The loop lives in the prompt, not in runtime code; every worker gets the same prompt except its ID.

Three 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 labeled OBSERVED, FACT, FAIL, CLAIM and PATCH_SUMMARY. Agensh sits on top of a single-agent harness; the experiments used GitHub Copilot CLI, and the paper says thin adapters can connect other harnesses such as Claude Code.

On the five hardest ProgramBench tasks—FFmpeg, gromacs, pandoc, PHP-src and ctags—agents rebuild each program from observed behavior, offline and within six hours. Reference repositories range 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 rose from 19.31% with 1 agent to 20.68% with 8, 26.52% with 32 and 28.78% with 128—a 9.47-point gain, or about 49% relative, including a 1.37-point gain from 1 to 8 agents. More agents raised scores on all five tasks and reached them sooner.

On pandoc alone, scores were 33.89% with 1 agent, 41.75% with 8, 43.11% with 32, 50.94% with 128 and 55.06% with 1,024. Speed improved as well: with 128 agents, pandoc passed 30% at the 30-minute checkpoint; 32 and 8 agents needed 60 and 90 minutes. The README adds a newer model: with GPT-6.1-sol, pandoc rose from 56.19% with 1 agent to 79.59% with 32 agents, a 41.64% relative gain. Those GPT-6.1-sol figures are not in the v1 paper.

The team traced worker trajectories at each scale. At 8 agents, peers agreed on interfaces and resolved overlapping claims themselves. At 32, peers reviewed shared work, and one approval was withdrawn after a counterexample. At 128, workers reused trusted reviewers and standardized an integration protocol. At 1,024, several workers acted as integrators, and peers recovered failed work. None of these roles was hard-coded. The release includes a single-node Docker Compose team and a single-agent baseline, both using native Copilot ACP sessions; it requires Linux with Docker Engine, Compose v2, systemd, cgroup v2 and Python 3.11+. Defaults include 8 workers on a single node; the 1,024-agent run used 16 nodes with 64 agents each, and workers have a 64 GiB memory ceiling. Each worker gets a 272,000-token input and 128,000-token output limit.

The paper compares Agensh with other multi-agent harnesses. Agensh has no orchestrator and peers self-claim work; Claude Code agent teams use a fixed lead session to coordinate teammates; Kimi Agent Swarm uses an orchestrator to decompose tasks; Copilot CLI /fleet assigns parts to subagents. Agensh was tested with 1,024 agents and has no hard limit, with 3 to 5 agents recommended; Kimi lists up to 100 sub-agents. Agensh is model-agnostic and was tested with GPT-5.6-sol and GPT-6.1-sol. It is open source under MIT. Claude Code agent teams are in Claude Code, experimental and off by default; Kimi Agent Swarm is a research preview for top-tier subscribers; Copilot fleet is in Copilot CLI.

Editor's Summary

Microsoft Research has released Agensh, an open-source multi-agent harness that removes the central orchestrator and lets coding agents self-claim tasks and merge work through shared Git and messaging services. Tests on ProgramBench from 1 to 1,024 agents showed higher scores and faster results as agent counts rose, though the five-task mean remained below 30% at 128 agents with GPT-5.6-sol. The release is model-agnostic, MIT-licensed and runs on Linux with Docker Compose.