Microsoft Open-Sources Agensh Multi-Agent Harness

Unlike most agent frameworks that rely on a central orchestrator, Agensh allows each worker agent to independently select its own sub-task, share its findings, and merge code into a common Git repository.

The Agensh multi-agent harness was tested with setups ranging from 1 to 1,024 agents. Each worker agent has an input limit of 272,000 tokens and an output limit of 128,000 tokens. The system runs on Linux with Docker Compose, defaulting to a single node with 8 workers, though the 1,024-agent configuration utilized 16 nodes, each running 64 agents. Workers are subject to a 64 GiB memory ceiling.

Performance evaluations on the ProgramBench benchmark indicated that increasing the number of agents improved scores on all five of the hardest tasks and accelerated their completion. Specifically, a 79.59% pandoc pass rate was achieved with 32 agents on GPT-6.1-sol, while GPT-5.6-sol reached 55.06% with 1,024 agents. However, the mean score across the five tasks only reached 28.78% at 128 agents. The transition from 1 to 8 agents resulted in a modest 1.37-point increase.

Agensh addresses the limitations of orchestrator-worker designs, where a lead agent handles all planning, assignment, and merging. Agensh enables workers to manage this coordination themselves through a concurrent and asynchronous 5-step loop. This loop involves gathering context, claiming a sub-task, taking action with tools, verifying results against acceptance criteria, and merging progress into a shared repository.