Agensh: Microsoft scales boss-free agent organizations to 1,024 coding agents
rohanpaul_ai · x · 2026-10-05
The arXiv paper "Agensh: Scaling Organizational Intelligence to 1,024 Agents" (Microsoft; authors include Li Dong and Furu Wei) drops the central orchestrator from multi-agent coding harnesses.
Workers run an asynchronous cooperation loop: continuously gathering context, claiming and self-assigning sub-tasks, acting, sharing findings, verifying, and merging progress. The infrastructure has three parts: a shared workspace for proposed/ongoing/completed work, a message interface between workers, and shared context holding reusable findings and intentions.
Evaluated with GPT-5.6-sol (high) on the 5 hardest ProgramBench tasks: scaling 1→128 agents raised the mean test-pass rate from 19.31% to 28.78% (49% relative), with larger organizations reaching comparable scores sooner; on pandoc, 1,024 agents lifted the pass rate from 33.89% to 55.06%. Worker trajectories show emergent self-organized cooperation.
Related event: Microsoft's Agensh Scales Decentralized Coding Agents to 1,024(2 posts)→
More from coding & agent
- SuperX MCP in Claude Code grew an X account from 64 to 1,000+ followers since July — tibo_maker · 2026-10-05
- Vanishing Gradients podcast: agents are like coffee — they help you make stupid mistakes faster — hugobowne · 2026-10-05
- Codex resets only unlock near your limit — tip: burn a bit of usage first to redeem expiring resets — JeremyNguyenPhD · 2026-10-05
- A permit layer for MCP tool calls makes same-key retries return the original receipt — HotPocketWaves · 2026-10-05
- AI agent produces a 1:45 video end-to-end on a local PC with Pixio and FFmpeg — tsi_org · 2026-10-05
- Stripe and Tempo Launch MPP, a New Payment Protocol Built for AI Agents — glenbeer · 2026-10-05