Microsoft open-sources Agensh, a harness for 1,024 coding agents

Microsoft Research has open-sourced Agensh, a decentralized harness that coordinates large teams of coding agents without a central orchestrator. On ProgramBench, scaling from 1 to 1,024 agents lifted the mean score on the five hardest tasks from 19.31% to 28.78%, and GPT-6.1-sol with 32 agents reached 79.59% on pandoc.

• 1,024 agents ran on 16 nodes with 64 workers each and a 64 GiB memory ceiling
• On pandoc, 32 agents with GPT-6.1-sol hit 79.59% versus 56.19% for a single agent
• The five-task ProgramBench mean rose from 19.31% to 28.78% at 128 agents
• Code is MIT-licensed and runs on Linux via Docker Compose; token cost is undisclosed