Skip to content
Neural Daily

ResearchNews

Agensh: 1,024 AI agents working together, no boss

🇭🇺 Olvasd el magyarul →

Microsoft researchers open-sourced Agensh, a harness where up to 1,024 coding agents split up the work on their own. More agents deliver better results in the same time, but the gains are far from linear.

By: Mikola Máté Patrik · · 2 min read

Agensh replay dashboard titled How 1,024 agents cooperate, showing agent activity across 16 servers, the Gitea repository tree, Mattermost messages and shared context entries
Image: Microsoft Research · Screenshot · Frame from the official Agensh replay animation in the GitHub repository · Source

What launched?

On October 24, Microsoft researchers open-sourced Agensh, a multi-agent harness, under the MIT license (GitHub). The preprint appeared on arXiv on September 22 (arXiv). There is no central orchestrator: agents discover and claim sub-tasks, share findings, verify each other's results and merge progress asynchronously.

  • Shared workspace: a Gitea repository.
  • Messaging: Mattermost, with direct messages between agents and shared channels.
  • Shared context: a mechanism adapted from the DeLM project, where agents post short records such as claim, fail or patch summary.
  • Each agent's own loop runs on the GitHub Copilot CLI. The release includes a Docker Compose single-node setup and a reproduction guide.

What they measured

The team tested it on the five hardest ProgramBench tasks: agents had to rebuild reference programs from their behavior and bundled documentation, with no internet access and a six-hour budget. With GPT-5.6-sol, the mean test-pass rate went from 19.31% with 1 agent to 28.78% with 128 agents, about 49% relative improvement.

On the Pandoc document converter, 1 agent reached 33.89%, 128 agents 50.94% and 1,024 agents 55.06% in the same time. With the newer GPT-6.1-sol, the Pandoc result rose from 56.19% with 1 agent to 79.59% with 32 agents. The researchers say roles, integration workflows and coordination habits emerged on their own as the team grew.

Agensh chart of Pandoc test-pass rate versus number of agents: GPT-5.6-sol from 33.89% with 1 agent to 55.06% with 1,024 agents, GPT-6.1-sol from 56.19% to 79.59% with 32 agents
Microsoft Research · Press kit · Official results chart from the Agensh GitHub repository · Source

Source: Microsoft Research. Pandoc test-pass rate by number of agents with GPT-5.6-sol and GPT-6.1-sol.

Why it matters

Most multi-agent systems today rely on a controller agent that becomes a bottleneck when it has to direct many workers. Agensh shows that the tools human dev teams already use (repo, chat, shared notes) work for agents too. The numbers also show the limits: on Pandoc, 1,024 agents beat a single agent by only about 21 percentage points, at many times the compute cost. In practice a better model may deliver more than many parallel agents, so weigh this against your own budget.

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles