Microsoft Research open-sourced Agensh on October 24, 2026, a harness that lets as many as 1,024 AI coding agents work on one software project with no lead agent telling them what to do. For small development teams, the headline number is less important than the design idea behind it: coordination, not headcount, is usually what limits a team of AI agents.
What happened
Most multi-agent coding setups follow an orchestrator-worker pattern. One lead agent plans the work, hands out sub-tasks, and merges the results. Microsoft’s researchers argue that the lead agent becomes the bottleneck as the team grows, so Agensh removes it entirely.
In Agensh, every worker runs the same five-step loop: gather context, claim a sub-task, take action, verify the result against the sub-task’s acceptance criteria, and merge progress into a shared repository. Workers coordinate through three shared services: Gitea hosts the shared workspace with branches and pull requests, Mattermost carries a team channel plus direct messages for resolving overlapping claims, and a shared context board stores typed notes such as observations, failures, claims, and patch summaries. The cooperation loop lives in the worker prompt rather than in runtime code, and the experiments run on top of GitHub Copilot’s agent harness, with the paper noting that thin adapters could connect other harnesses such as Claude Code.
The team evaluated Agensh on the five hardest ProgramBench tasks — FFmpeg, gromacs, pandoc, PHP-src, and ctags — in which agents rebuild a program from its observed behaviour and bundled documentation, working offline within a six-hour budget. With GPT-5.6-sol at high reasoning effort, the mean test-pass rate across the five tasks rose from 19.31% with one agent to 28.78% with 128 agents, a relative improvement of about 49%. On pandoc alone, the pass rate climbed from 33.89% with one agent to 50.94% with 128 agents and 55.06% with 1,024 agents. Speed improved too: on pandoc, 128 agents passed 30% of tests within 30 minutes, while 32 agents needed 60 minutes and 8 agents needed 90 minutes to reach the same mark. With the newer GPT-6.1-sol, the pandoc pass rate rose from 56.19% with one agent to 79.59% with 32 agents, according to the project README (these figures are not in the version 1 paper).
As teams grew, behaviours the researchers did not hard-code began to emerge. Eight agents agreed on interfaces and resolved overlapping claims themselves; 32 agents reviewed one another’s work; 128 agents reused trusted reviewers and standardised an integration protocol; and at 1,024 agents, several workers acted as integrators while peers recovered failed work.
Why it matters for small businesses
No small business needs 1,024 coding agents, and that is not the point. The useful takeaway is architectural. If your agency or in-house team already uses coding agents and hits a wall when several agents work on the same codebase — conflicting edits, duplicated work, or a lead agent that becomes a bottleneck — Agensh is a reference design for an alternative: let workers claim small tasks, publish findings early, and merge through the same Git workflow your human developers already use.
It also sets realistic expectations. Even with 128 agents, the mean pass rate on these deliberately hard rebuild tasks was under 30%. More agents helped, but they did not turn a hard task into an easy one, and the gains flattened as teams grew. For everyday business code — bug fixes, small features, test coverage — a handful of well-coordinated agents with human review is a far more sensible starting point than a swarm.
Pricing, availability, and what to do next
Agensh is open source under the MIT licence, and the release includes a single-node Docker Compose team setup (eight workers by default) plus a single-agent baseline, both using native Copilot ACP sessions. You will need Linux with Docker Engine, Compose v2, systemd, cgroup v2, and Python 3.11 or later, plus ProgramBench version 1.2.5 and a model API token. Note that token costs are not disclosed in the release, so budget for model usage separately — the code is free, but the model calls are not.
If you want to try it, start small: run the single-node setup against a contained, non-production repository, compare its output against your current single-agent workflow on the same task, and keep a human in the merge loop. Treat the 1,024-agent result as proof that leaderless coordination can scale, not as a target to copy.
FAQ
Is Agensh a product I can buy? No. It is an open-source research harness from Microsoft Research, released under the MIT licence, not a commercial product with pricing or support tiers.
Do I need 1,024 agents to use it? No. The release ships a single-node setup that runs eight workers by default, which is the sensible starting point for a small team.
Does Agensh replace the coding agent I already use? No. It sits on top of a single-agent harness — the experiments used GitHub Copilot’s harness — and coordinates many instances of it.
Will more agents always produce better code? Not necessarily. On the benchmark, scores improved with team size but with diminishing returns, and even large teams scored under 30% on average across the hardest tasks.
Sources
Microsoft Agensh repository (GitHub): open-source announcement dated October 24, 2026, MIT licence, benchmark figures and reproduction guide.
MarkTechPost: Microsoft Open-Sources Agensh: A Multi-Agent Harness for 1,024 Agents, published October 25, 2026, with the full comparison table and setup requirements.
