Microsoft Research has released Agensh as open source, a framework that lets large groups of AI coding agents work together. Its distinguishing feature is the absence of a central coordinator: in most existing systems a lead agent plans, assigns and merges the work of all the others. In Agensh, that role does not exist, and each agent decides for itself which part of the project to tackle.
Every agent runs the same loop, in parallel and asynchronously. It reads the shared goal, its peers' progress and any messages received, then posts a claim declaring the task it intends to take on. If two agents claim overlapping work, they message each other directly to adjust the split. It then works with its tools, checks the result against the task's acceptance criteria, merges the code into the shared repository, posts an update and starts again.
Image: microsoft — https://github.com/microsoft/Agensh
Three services make this possible. A Gitea server hosts the common repository, with branches, pull requests and conflicts. Mattermost provides a team channel and direct messages for urgent collisions. A shared context board, adapted from an earlier project, stores typed notes such as observations, facts, failures, claims and change summaries. The coordination logic lives in the prompt rather than in the system's code: every agent receives the same prompt, changing only its identifier.
Tests were run on ProgramBench, a benchmark that asks for programs to be rebuilt from their observed behaviour, working offline within six hours. The five hardest tasks were used, including FFmpeg, gromacs, pandoc, PHP-src and ctags. With the GPT-5.6-sol model, the mean share of passed tests rose from 19.31% with one agent to 28.78% with 128 agents, a relative gain of about 49%. On pandoc alone the result went from 33.89% with one agent to 55.06% with 1,024 agents. Speed improved too: with 128 agents pandoc passed 30% of its tests after 30 minutes, while 8 and 32 agents needed 90 and 60 minutes.
With the newer GPT-6.1-sol model, pandoc went from 56.19% with one agent to 79.59% with 32 agents. These figures do not appear in the first version of the paper that accompanies the release.
As the number of agents grew, the research team observed behaviours nobody had programmed. With eight agents, peers agreed on interfaces and settled overlaps themselves. With 32, they reviewed each other's work, and one approval was withdrawn after a counterexample. With 128, agents reused trusted reviewers and standardised an integration protocol. With 1,024, some agents took on integrator roles and others recovered failed work.
Agensh is distributed under the MIT licence and installs on Linux with Docker Compose, with a default setup of eight agents per node. The 1,024-agent run used sixteen nodes, with sixty-four agents each and a 64 GiB memory limit per agent. Microsoft has not disclosed the token cost of the run, a figure that remains decisive for judging how well the system scales.
No comments yet. Be the first!