Microsoft opens a tool that lets 1,024 AI coding agents work together without a boss

What you could do with it

• A researcher could test how AI coding agents cooperate by scaling a team from one agent up to 1,024 agents rebuilding real open-source programs like Pandoc and FFmpeg ¹.
• A developer could let dozens of AI agents coordinate on their own to tackle bigger programming tasks and build software faster ¹.
• A researcher could study how useful roles like integrators and reviewers emerge naturally among workers without being programmed in ¹.

A team with no manager

Most multi-agent systems use a lead agent that plans work, assigns tasks, and combines results — like a project manager. Agensh removes that central boss entirely. Each worker agent independently picks its own sub-task, shares what it finds, and merges code into a shared Git repository ¹.

The system was tested on [ProgramBench], a benchmark where agents rebuild real open-source programs like Pandoc, FFmpeg, and PHP from scratch based on how they behave. The Microsoft team scaled the setup from just one agent up to 1,024 agents simultaneously ¹.

How the agents cooperate

Every worker follows the same five-step loop, running at the same time as all the others: it gathers context about the goal and what peers are doing, claims a sub-task by posting in a shared channel, works on it using its tools, checks the result against acceptance criteria, and merges progress into the shared code repo before repeating ¹.

Three simple services connect the workers: [Gitea] hosts the shared workspace with branches and code reviews, [Mattermost] carries team messages and direct chats for urgent collisions, and a shared context board stores typed notes: OBSERVED, FACT, FAIL, CLAIM and PATCH_SUMMARY ¹.

The coordination lives inside the prompt given to each agent, not in special runtime code. The experiments used GitHub Copilot CLI as the base harness, though the researchers say other systems like Claude Code could be connected with thin adapters ¹. Each worker receives a 272,000-token input and has a 128,000-token output limit .

Big gains at scale

On the five hardest ProgramBench tasks, adding more agents steadily improved results ¹. Using GPT-5.6 Sol at high reasoning effort, the average pass rate across all five tasks rose from 19.31% with one agent to 28.78% with 128 agents — a gain of about 49% relative .

On the Pandoc task specifically, scores climbed from 33.89% (one agent) to 55.06% (1,024 agents) ¹. Speed also improved: with 128 agents, 30% of Pandoc tests passed within 30 minutes, while 32 and 8 agents needed 60 and 90 minutes respectively to reach the same checkpoint .

A newer model, GPT-6.1 Sol, pushed the Pandoc score to 79.59% with just 32 agents — a 41.64% relative improvement over a single agent ¹. These GPT-6.1 figures are not included in the original research paper .

Roles emerged on their own

The researchers tracked how agents behaved as the team grew and found that useful roles formed naturally without being programmed in ¹. With eight agents, peers agreed on interfaces and resolved overlapping claims through direct messages. At 32 agents, workers began reviewing each other's shared work. At 128 agents, they standardized an integration protocol and reused trusted reviewers. At 1,024 agents, some workers acted as integrators while others recovered failed work .

Running Agensh yourself

The release includes a single-node setup running eight workers by default and a single-agent baseline, both using Docker Compose on Linux ¹. You need Docker Engine, Compose v2, systemd, cgroup v2, and Python 3.11 or later, plus a ProgramBench installation and a model API token . Full instructions are in the reproduction guide . The code is released under the MIT license, though the researchers did not disclose token costs .

How it compares

Data from the Microsoft Research release ¹.

This story reports on a trend for information only. It is not advice or encouragement to do any of it.

How good is it

Microsoft researchers tested Agensh on ProgramBench, where agents rebuild real open-source programs from scratch ¹.,With GPT-5.6 Sol at high reasoning effort, the average pass rate across the five hardest tasks rose from 19.31% with one agent to 28.78% with 128 agents, a gain of about 49% relative .,On the Pandoc task specifically, scores climbed from 33.89% with one agent to 55.06% with 1,024 agents .,Speed also improved: 128 agents passed 30% of Pandoc tests within 30 minutes, while 32 and 8 agents needed 60 and 90 minutes to reach the same checkpoint .

Key facts

• Microsoft open-sourced Agensh, a multi-agent system for up to 1,024 coding agents with no central coordinator
• On ProgramBench, 128 agents lifted the average pass rate from 19.31% to 28.78% with GPT-5.6 Sol
• GPT-6.1 Sol with 32 agents reached 79.59% on the Pandoc task, a 41.64% improvement over one agent
• Agensh runs on Linux with Docker Compose, using Gitea, Mattermost, and a shared context board
• The code is MIT licensed; token costs were not disclosed

Sources

1. Microsoft Open-Sources Agensh: A Multi-Agent Harness for 1,024 Agents — marktechpost.com