Agensh coordinates up to 1,024 coding agents without a central orchestrator

Microsoft Research’s Agensh is a self-organized multi-agent harness with no central orchestrator. Parallel coding agents coordinate through shared state: each gathers context, claims a sub-task, does the work, shares what it found, verifies the result, and merges it asynchronously through a shared workspace and a message channel. The setup was tested with more than 1,000 coding agents running at once.

On the five hardest ProgramBench tasks with GPT-5.6-sol, increasing the agent count from 1 to 128 raises the mean final test-pass rate from 19.31% to 28.78%. Larger teams also reach a given pass rate sooner. On pandoc, 1,024 agents take the test-pass rate from 33.89% to 55.06%. The authors also report forms of cooperation that the agents start on their own and that become standard practice as the team grows.

The paper is available on arXiv.