Agensh

Agensh is an open-source multi-agent harness from Microsoft Research in which concurrent agent workers gather context, claim sub-tasks, share findings, verify results and merge progress asynchronously, with no central orchestrator. The project's authors report evaluations scaling from 1 to 1,024 agents on ProgramBench tasks. The code is MIT-licensed, and the latest listed release is v1.1.

What It Is

Agensh is a harness that defines a cooperation loop for many agents working on one shared goal, such as building out a software repository. Each worker proceeds independently and shares progress through a lightweight agentic organization infrastructure made up of a shared workspace, a message interface and shared context. Gitea provides the shared workspace, Mattermost provides messaging, and the shared context mechanism is adapted from DeLM.

Backends and Setup

Agensh is layered on top of single-agent harnesses. Copilot is the default and was used for the 1.0 experiments, and Pi, Codex and Claude Code are also supported, selected with an --agent-harness option. The release includes a single-node team setup using Docker Compose and a standalone single-agent baseline. Users can also run a team on their own software repository, and the documentation includes a reproduction guide, harness configuration and launch scripts.

Reported Scaling Results

The authors state that, with GPT-5.6-sol at high reasoning effort, a six-hour budget and no Internet access, mean final test-pass rate across the five hardest ProgramBench tasks rose from 19.31% with 1 agent to 28.78% with 128 agents. On Pandoc, they report 33.89% with 1 agent and 55.06% with 1,024 agents. They also report that peer coordination, integration workflows and specialized roles emerge as the organization grows. A replay demo of the 1,024-agent run is published on the project page.