Microsoft's Agensh: boss-free coding agent teams keep improving from 1 to 128 agents
rohanpaul_ai · x · 2026-10-05
A new Microsoft paper introduces Agensh, a self-organized multi-agent coding harness with no central orchestrator: each agent claims its own sub-task, builds and tests it, merges into a shared Git repo, and logs findings on a shared board.
On the 5 hardest of ProgramBench's 200 tasks (using GPT-5.6-sol high), scaling from 1 to 128 agents raised the mean test-pass rate from 19.31% to 28.78% (49% relative gain), improving at every step, and larger teams reached comparable scores sooner. On pandoc, scaling to 1,024 agents pushed the pass rate from 33.89% to 55.06%.
Popular multi-agent coding tools funnel all work through a single lead agent with limited bandwidth. The paper used a single model throughout and doesn't report the cost of large teams.
Related event: Microsoft's Agensh Scales Decentralized Coding Agents to 1,024(2 posts)→
More from coding & agent
- RAEA engine lets AI agents create videos, 3D scenes and games with React — Promptmethus · 2026-11-22
- Sentry CEO: keep design docs in the repo — agents need the intent — zeeg · 2026-11-22
- Antigravity team digests 500+ user feedback items; cross-workspace permission friction flagged — dotey · 2026-11-22
- kvdiff (MIT): a 63kB local proxy that byte-pinpoints MCP-induced prompt cache busts and auto-heals them — BringinItDirty · 2026-11-22
- Open-source quantified-self MCP server queries Apple Health, WHOOP and Oura locally in Claude Desktop — thecimal · 2026-11-22
- Alchemy Run's distilled SDK now beats first-party AWS SDK v3 on load time — samgoodwin89 · 2026-11-22