Microsoft open-sources Agensh for 1,024 agents at one time
The agents of Agensh share a code repository, a chat and a findings board, with no central controller.
Microsoft Research released Agensh as open source. Agensh is control software that operates many AI agents on one task. It has no central controller to give tasks to other agents. Microsoft says the agents organize the work between them. It says they scale from 1 to 1,024 agents on ProgramBench, with a limit of 6 hours.
How it works
The agents organize the work between them. They share a Git repository (a record of code versions), a chat and a findings board. No one agent controls the group.
The numbers
The source gives one range: 1 to 1,024 agents on ProgramBench, with a 6-hour limit. It gives no scores, no cost and no test against other tools.
What is not known
The source gives no information about ProgramBench or the 6-hour limit. It does not show how results change with more agents.
Insight: Microsoft builds open tools for each step of agent work
On 20 October, Microsoft released Scope, a tool that tests coding agents. Agensh operates many agents together. A group that includes Microsoft also gave OpenEnv, a tool for training worlds, to the Agentic AI Foundation. This reads as Microsoft placing open tools at the test, operate and training steps, possibly to set a standard. Monitor if Agensh joins a neutral group.
Sources
Posted
