Microsoft’s repository says Agensh, its experiment launch scripts, and a reproduction guide were open-sourced on October 24, 2026, under the MIT License. Agensh coordinates coding agents without a central orchestrator. Its reported 1,024-agent result is for the Pandoc task; the five-task average was reported through 128 agents.
What Agensh does
A harness is software that organizes how AI agents work on a task. In Agensh, workers gather context, claim subtasks, take action, verify results, and merge progress. They share work asynchronously rather than waiting for one central orchestrator, a lead system that assigns tasks and combines results.
The project describes Gitea as the shared workspace and Mattermost as the messaging interface. Agents also use shared context adapted from DeLM to retain useful findings and work intentions. This gives workers common places to exchange code, messages, and context as they tackle subtasks.
What the reported results measure
The Agensh paper reports tests on five ProgramBench software-reconstruction tasks using GPT-5.6-sol at high reasoning effort, with a six-hour budget and no internet access. Its five-task mean final test-pass rate rose from 19.31% with one agent to 28.78% with 128 agents. The paper reports a separate Pandoc result of 33.89% with one agent and 55.06% with 1,024 agents.
| Evaluation | Model | Agent counts compared | Final test-pass rates | Conditions |
| Five-task ProgramBench mean | GPT-5.6-sol | 1 and 128 | 19.31% and 28.78% | High reasoning; six hours; no internet |
| Pandoc | GPT-5.6-sol | 1 and 1,024 | 33.89% and 55.06% | High reasoning; six hours; no internet |
| Pandoc | GPT-6.1-sol | 1 and 32 | 56.19% and 79.59% | High reasoning; six hours |
Microsoft’s repository separately reports the GPT-6.1-sol Pandoc figures. The 1,024-agent number therefore belongs to one task-specific result, not the five-task mean. Together, these figures describe performance in the stated evaluations; they do not rank Agensh against other harnesses or establish how it performs across coding tasks generally. The Agensh paper and Microsoft’s project repository provide the original research and project details.
What the open-source release includes
The repository provides a single-node Docker Compose setup for a team of agents, a standalone single-agent baseline, experiment launch scripts, and a reproduction guide. The MIT license applies to the project’s code.
