Microsoft Research has open-sourced Agensh, a system that coordinates teams of AI agents for coding without a central orchestrator. Instead of a lead agent handing out tasks, each agent claims its own work, shares findings and merges progress into a shared repository.
Each agent runs the same loop, concurrently and asynchronously: gather context, claim a sub-task, act and share findings, verify the result and merge progress. Three components support the loop: a shared workspace, a message interface and a shared context board with reusable notes.
On the five hardest ProgramBench tasks with GPT-5.6-sol, the mean pass rate rose from 19.31% with 1 agent to 28.78% with 128 agents. Rebuilding pandoc from scratch in 6 hours without internet, the rate went from 33.89% with 1 agent to 55.06% with 1,024 agents. With GPT-6.1-sol, the project README reports 79.59% on pandoc with 32 agents — a newer figure outside the paper's first version.
More agents are not always better: gains shrink as teams grow, and token costs were not disclosed. Agensh runs on Linux with Docker Compose, builds on GitHub Copilot CLI for each agent, and needs a model API token and ProgramBench to reproduce. It fits complex tasks under tight deadlines, not simple questions.
Image credit: first page of the Agensh paper, via DAIR.AI Academy.
Sources: official Agensh project page; MarkTechPost; DAIR.AI Academy.
Transparency: this article was produced with the help of artificial intelligence, cross-checked against public posts on X and internet sources. Check the original sources before sharing or making decisions based on it.
Muse
