Microsoft releases Agensh for self-organizing coding agents

Microsoft’s Agensh release, announced on October 24, makes its agent harness and experiment scripts available. A harness supplies the working environment around AI agents. Here, workers choose and claim smaller tasks, share findings, check results and merge progress asynchronously, rather than waiting for a central coordinator. Gitea holds their shared workspace, Mattermost carries messages, and shared context connects individual progress to the team.

The shared-context service stores admitted findings and work claims in MySQL, a database, while a recent in-memory view supplies information to prompts. Its checks examine text deterministically: they do not execute the verification commands a worker reports. A note being accepted therefore does not establish that its claimed test ran. By default, claims disappear from live views after an hour, while persisted history remains searchable.

For developers considering an experiment, the reproduction guide specifies Linux and Docker infrastructure. It also documents consequential limits: agents share peer roles, pull requests can merge without required approvals, and the continuous-integration merge gate is disabled. The released setup uses a different cleanroom reference from the historical experiment. The guide consequently asks users to record their own release revision instead of claiming exact historical reproduction.

MarkTechPost covered the release; its article summarizes the research and repository.