Microsoft opens Agensh code for agent teams that choose their own work
Microsoft has released the Agensh harness, experiment launch scripts and reproduction guide for software teams operating without a central orchestrator. Workers choose subtasks, exchange findings and merge verified changes. A shared code workspace, messaging and shared context connect their progress. The code opening on 24 October is a separate step from the research preprint published in September.
Artificial Intelligence··Night
Code joins the published research
Microsoft has opened the code for Agensh, a harness that lets coding agents organise their own teamwork. The repository dates the release of its implementation, launch scripts and reproduction guide to 24 October. AI Tool Herald also reports the code opening and the absence of a central orchestrator assigning every step.[1], [2]
Workers choose and merge tasks
Agensh workers gather context, claim tasks, make changes, verify their work and merge it asynchronously. Each worker runs through its own Copilot agent harness. The shared workflow leaves task selection to the workers, with their findings and completed changes feeding back into the common workspace while other workers continue.[1]
The implementation uses Gitea for the shared code workspace and Mattermost for messages. Shared context is adapted from DeLM. These components give workers places to exchange findings and coordinate changes without replacing them with a single manager that decomposes the whole job in advance.[1]
A runnable package and a baseline
The release includes a single-node Docker Compose setup and a standalone single-agent baseline using the native Copilot ACP interface. The reproduction guide identifies entry points and the repository layout. The code carries an MIT licence. This package follows the research preprint published on 22 September, supplying implementation and run instructions as a later release step.[1]