Microsoft Research has open-sourced Agensh, a decentralized harness designed to coordinate large teams of AI coding agents without relying on a central orchestrator. Traditional agent architectures depend on a primary lead model to plan and assign sub-tasks, creating performance bottlenecks. Agensh replaces this model by enabling each worker agent to autonomously select sub-tasks, communicate via messaging channels, and merge code into a shared Git repository.

In benchmarks on ProgramBench using GPT-5.6-sol, scaling from 1 agent to 1,024 agents improved the mean test-pass rate from 19.34% to 28.81%. The system also demonstrated faster issue resolution, with 128 agents reaching target progress metrics in one-third of the time required by smaller teams.

The open-source release includes a Docker Compose environment and supports integration with various single-agent frameworks such as GitHub Copilot CLI and Claude Code.

Why it matters

  • Removes orchestrator performance ceilings, enabling massive parallelization for complex enterprise software engineering tasks.

  • Provides actionable architecture blueprints for operating multi-agent swarms using existing version control and communication tools.

  • Accelerates execution speed for large codebase refactoring and autonomous bug fixing.

Source: marktechpost.com