The Pulse
Microsoft Releases Agensh, Tested With 1,024 AI Agents
Microsoft open-sourced Agensh on October 24, a multi-agent harness designed to let AI workers coordinate without a central orchestrator. In tests reported by its authors, 1,024 agents raised Pandoc’s six-hour test-pass rate to 55.06%, thoug
AI.info Team · Written with AI · fact-checked ·
“Scaling from 1 to 128 agents raises the mean final test-pass rate from 19.31% to 28.78%, an approximately 49% relative improvement.”
Zhihao Zhan et al., Microsoft Research
Microsoft made Agensh public on October 24, releasing the multi-agent harness, experiment launch scripts and a guide for reproducing its tests. The project is designed to let software-building agents divide work and combine results without a central agent assigning tasks. Its authors had introduced the system in a September 22 preprint; the new GitHub repository publishes the implementation under an MIT license.
Workers claim tasks and merge their work
Agensh gives each worker access to three shared coordination tools: a workspace for proposed and completed work, a messaging system, and shared context for findings and active claims. Workers read peer updates, claim tasks, act, check their results and merge contributions, then repeat the cycle. They do so asynchronously, so one worker need not wait for the rest of the team to finish a step.
The reported implementation uses Gitea for the shared workspace and Mattermost for messages, with shared context adapted from DeLM. Agensh sits above an individual agent’s existing runtime rather than replacing it; Microsoft’s experiments use the Copilot harness. The paper describes lightweight adapters as a way to connect the organization layer to other underlying harnesses, including Claude Code.
The 1,024-agent result comes from Pandoc
The authors tested the system on ProgramBench, where agents must rebuild software from observed behavior and bundled documentation without internet access. Across five difficult tasks—FFmpeg, GROMACS, Pandoc, PHP-src and ctags—the mean final test-pass rate rose from 19.31% with one agent to 28.78% with 128, under a six-hour limit. The model was GPT-5.6-sol with high reasoning effort.
The 1,024-agent figure comes from a separate Pandoc test, not the five-task average. There, the final pass rate went from 33.89% for one agent to 50.94% for 128 and 55.06% for 1,024, after six hours. The paper says the 1,024 workers ran across 16 machines, with 64 agents per machine, and that activation was staggered to reduce clashes over work.
The comparison has a prompt-level caveat
The headline scores are the authors’ experimental results, not proof that adding workers alone explains every gain. Their report says the one-agent baseline received a special stop-hook prompt because single-agent runs rarely sustain the full six hours; the multi-agent organization instead used an idle detector that prompts workers after ten minutes without activity. That difference means readers should treat the reported comparison as a result for the tested setups, rather than a clean isolation of agent count.
The 1,024-agent result also measures one task, Pandoc, while the five-task average stops at 128 agents. The authors report no internet access during the rebuilds, which rules out looking up the target projects’ existing implementations but limits what the tests establish about work outside this benchmark. The release and paper do not state an inference-cost figure.
Reproducing the setup takes more than cloning the repo
The repository includes a single-node Docker Compose team setup and a standalone single-agent baseline, both using native Copilot ACP sessions. Its reproduction guide calls for Linux, Docker Engine, Compose v2, systemd and cgroup v2, along with a configured model endpoint and credentials. The documented defaults are eight workers and six hours per task.
That makes the release useful to developers who want to inspect the harness or rerun its benchmark, while the paper’s largest result remains a bounded test of parallel software reconstruction. For Pandoc, the measured difference between 128 and 1,024 agents was 4.12 percentage points over the same six-hour budget.