Overview
Agensh is a multi-agent coding framework open sourced by Microsoft Research and announced on October 24, 2026. The problem it targets is specific: once dozens or hundreds of coding agents run in parallel, a central orchestrator becomes the bottleneck, since assignment, progress and conflict handling all squeeze through one process. Agensh removes that point, letting each worker claim subtasks from a shared board, publish findings in the same place and merge changes straight into a shared Git repository.
Structurally it is three pieces: Gitea for hosting and merging code, Mattermost for communication and a shared task board tying them together. Workers do not relay messages through an orchestrator; board state and repository commits are the coordination signal. The paper reports that on the five hardest ProgramBench tasks, pass rates rose from 19.31 to 28.78 percent with 128 agents, and on the pandoc task scaling to 1,024 agents lifted pass rates from 33.89 to 55.06 percent.
Maturity deserves a caveat. The paper lists github.com/microsoft/Agensh as the repository, and it did not exist at submission time, so how much code and reproduction material is public should be read from the paper itself. It reads more as a technical route laid out for inspection: multi-agent coding need not depend on central scheduling, and scale itself can be a source of capability.
Key Features
- No central orchestrator: There is no scheduling process: workers claim subtasks from the shared board themselves. Assignment, progress and conflict handling all happen in shared media, so scaling up does not put pressure on a single process.
- Shared Git repository as the merge surface: Every worker merges into the same repository and conflicts resolve in Git. Code becomes the coordination signal, and the commit history records who saw what and changed what.
- Gitea plus Mattermost: Code hosting runs on Gitea and coordination on Mattermost, both self-hostable open source projects. The whole stack can sit inside a private network without external services.
- Scale buys pass rate: The paper measures the scaling curve on ProgramBench: 28.78 percent on the five hardest tasks with 128 agents, and 33.89 to 55.06 percent on pandoc at 1,024 agents. Gains are not linear, but difficult tasks do improve with scale.
- Workers claim their own work: Each worker decides whether to take a subtask, shares what it finds and then decides the next step, spreading planning cost across the fleet. It also means task granularity has to be designed carefully.
- Self-hostable stack: All three components are self-hosted, so the experiment can run fully enclosed. For teams that need large-scale coding experiments behind a firewall, this matters more than the benchmark numbers.
Use Cases
- A research environment for multi-agent scaling laws, changing agent count and task difficulty on the same harness.
- Parallel attempts on hard benchmarks, letting workers propose different solutions and filtering by test results.
- Batch coding work inside a private network, with Gitea and Mattermost local and code never leaving the firewall.
- Teams reproducing the paper, assembling the stack as described and comparing central orchestration against decentralization.
- Teaching multi-agent coordination, where a shared board and repository make the mechanics visible.
Pros
- Without a central orchestrator, scaling up does not create a single-process bottleneck.
- The shared Git repository provides merging, rollback and history for free, so conflict handling needs no new mechanism.
- Gitea and Mattermost are both self-hostable, so the whole stack can run inside a private network.
- The paper publishes scaling curves on ProgramBench and pandoc, so the return on scale can be checked against numbers.
- Workers claiming their own work simplifies the scheduling layer, leaving granularity to the task board.
- It leans on open source components, leaving room to modify or replace individual pieces.
Pricing
Agensh is an open source Microsoft Research project and the code is free. Running costs come from the number of agents and model calls you host: the framework itself is free, Gitea and Mattermost can be self-hosted, and model usage is billed by whichever API you choose. Experiments at 1,024 agents demand substantial compute and concurrency, so a smaller pilot before scaling up is sensible. Check the paper and the official repository for how much code and reproduction material is available.
Summary
Agensh is valuable for making one route through multi-agent coding explicit: drop the orchestrator, claim work from a shared board, merge into one Git repository, and scale upward. The paper scaling curves, from 19.31 to 28.78 percent on the hardest five tasks and 33.89 to 55.06 percent on pandoc, show the route holds. It remains a research project with limited code availability, but as a reference point against centrally scheduled designs it is worth reading for anyone working on multi-agent architecture.
Version History
- Microsoft Research open sources Agensh, a multi-agent coding harness with no orchestrator (2026-10-24): Microsoft Research announced Agensh on October 24, 2026, a multi-agent coding framework that removes the central orchestrator so workers claim subtasks from a shared board, publish findings and merge changes into one Git repository, with Gitea for hosting and Mattermost for coordination. The paper reports pass rates on the five hardest ProgramBench tasks rising from 19.31 to 28.78 percent with 128 agents, and 33.89 to 55.06 percent on pandoc at 1,024 agents. The repository listed in the paper, github.com/microsoft/Agensh, did not exist at submission time.