1024 AI Agents Without a Central Commander: Microsoft Open-Sources Agensh

Microsoft Research has officially open-sourced Agensh, a multi-agent runtime framework built around a very clear core design: no central orchestrator at all.

The standard common approach to multi-agent collaboration is to have a master scheduler that assigns tasks and collects returned results. This works perfectly well when you only run a small number of agents, but once the agent count scales up, the scheduler quickly becomes a major bottleneck, and state synchronization and exception handling also grow far more complicated. Agensh takes a completely different path: all worker agents self-organize, sharing a single Git repository, chat system, and findings board. Code and deliverables are stored in the shared repository, communications and discoveries are saved on these shared platforms, and collaboration status no longer needs to be forwarded through a single central node.

As reported by Asif Razzaq from MarkTechPost, Agensh can scale from 1 agent all the way to 1,024 agents on the ProgramBench benchmark, with a 6-hour time budget. This is a very intuitive number, but it also comes with clear boundaries: it is a result measured under specific evaluation conditions, which does not mean you can directly plug 1,024 agents into a production environment and expect it to run properly. Real-world projects still need to address permission control, merge conflicts, cost management and observability requirements.

The value this brings to developers is that it delivers a new engineering idea: sink coordination logic down to the shared repository and messaging layer, to cut the overhead of central scheduling significantly. For general readers, you can think of it as a group of workers who share the same codebase, group chat and result board, who find work on their own and verify completed outputs independently. Scaling agent count up to four digits sounds like moving an entire office meeting room onto servers, but at the very least, the team did not solve the scaling problem by adding an even busier central commander.

Agensh is already open sourced. If you want to try it out, start by reviewing the repository documentation and license agreement, reproduce the test results under the official ProgramBench conditions first, then evaluate how to connect it to your existing repositories and messaging systems. Original article link: Microsoft Open-Sources Agensh: A Multi-Agent Harness for 1,024 Agents.