- Microsoft Research open-sourced Agensh, a multi-agent coding harness that removes the central orchestrator most agent frameworks rely on.
- Workers pick their own sub-tasks, communicate through shared chat and a findings board, and merge code into a common Git repository.
- On the five hardest ProgramBench tasks, scaling from 1 to 128 agents raised the mean final test-pass rate from 19.31% to 28.78%, roughly a 49% relative gain.
- On the pandoc task specifically, scaling to 1,024 agents under a 6-hour budget pushed the pass rate from 33.89% to 55.06%.
What Happened
Microsoft Research open-sourced Agensh on October 24, 2026, after first posting the underlying paper to arXiv on September 22 and revising it October 25. The harness is designed to let large numbers of coding agents self-organize instead of waiting on a central planner to assign work.
Why It Matters
Most multi-agent coding frameworks route every decision through a single orchestrator, which becomes a bottleneck as the number of agents grows. Agensh tests whether removing that bottleneck lets more agents actually translate into better results, rather than just more parallel noise.
Technical Details
Each Agensh worker selects its own sub-task, shares progress through a chat channel and a findings board, and merges completed code into a shared Git repository without a coordinator approving the plan first. Researchers measured the approach on ProgramBench’s 200 coding instances, focusing on the five hardest: FFmpeg, gromacs, pandoc, PHP-src, and ctags. The pandoc task showed the clearest scaling effect, with the pass rate climbing from 33.89% at a single agent to 55.06% at 1,024 agents under the same six-hour time budget.
Who’s Affected
Teams building large-scale autonomous coding systems get a reference implementation for decentralized multi-agent coordination instead of designing one from scratch, and the ProgramBench results give a concrete benchmark for comparing orchestrator-free designs against traditional hub-and-spoke agent frameworks. The release lands alongside separate research on cutting multi-agent token costs and studies questioning whether bigger agent teams reliably improve quality, underscoring how unsettled multi-agent scaling still is.
What’s Next
Because the gains varied significantly by task, with pandoc scaling well but other ProgramBench targets likely to behave differently, the open release invites outside teams to test whether orchestrator-free scaling holds on workloads beyond the five tasks Microsoft highlighted.
