Microsoft open-sources Agensh multi-agent coding harness

Microsoft published the code and experiment reproduction tools for Agensh, its multi-agent harness for AI coding agents, on GitHub on October 24, according to aitimes.com. The release allows up to 1024 agents to work on tasks without a central manager.

Agensh does not assign work from above. Agents coordinate through a shared repository, a message system and a shared record, choosing their own tasks as they go. The harness adds collaboration features on top of existing agent harnesses and was built to connect to other harnesses as well.

Microsoft evaluated the system on ProgramBench across five difficult tasks, each with a six-hour limit and no internet access. The experiments used the GPT-5.6 Sol model and the GitHub Copilot CLI. With a single agent, the average final test pass rate was 19.31%; with 128 collaborating agents it rose to 28.78%, a gain of 9.47 percentage points and roughly 49% relative improvement.

Emergent roles at larger scale

In the pandoc reproduction task with 1024 agents, the final pass rate climbed from 33.89% for a single agent to 55.06%. In the same task, 128 agents passed 30% at the 30-minute mark, while a single agent did not reach that level in the first two hours.

At the 1024-agent scale, several agents took on code integration roles, and workers handed results to whichever responder was qualified first. All agents received identical instructions apart from their identifiers, so roles and procedures emerged from communication rather than being pre-assigned.

The related research paper was published earlier, in September 2026. According to the researchers, existing multi-agent systems mostly rely on a central agent that splits tasks and aggregates results, which they say limits performance as agent counts grow.