Microsoft open-sources Agensh, a decentralized multi-agent harness scaling to 1,024 agents
donglixp · x · 2026-10-24
Microsoft researchers released Agensh, an MIT-licensed, self-organizing decentralized multi-agent harness that scales concurrent cooperation from 1 to 1,024 agents, verified on building large software systems from scratch.
Key points:
- No central orchestrator: agents discover and claim sub-tasks, share findings, verify results, and merge progress asynchronously;
- A shared workspace, message interface, and shared context connect each agent for long-horizon coding tasks;
- A preprint was published on arXiv (2026-09-22), and the full harness, experiment launch scripts, and a reproduction guide are open-sourced on GitHub.
The repo includes the harness, demos, Docker setup, prompts, and a mini-router — reproducible end to end via REPRODUCE.md.
More from coding & agent
- After 3 Weeks of Struggling, Dev Says You Can't Vibe Code Everything — pramodk73 · 2026-10-24
- This developer routes 95% of tasks to DeepSeek and saves frontier models for the hardest 5% — mariofilhoml · 2026-10-24
- Harness engineering setup for Opus 5.5: model tiering, context hygiene and progress files that persist — Arindam_1729 · 2026-10-24
- He skipped a $12/mo embroidery app by having ChatGPT build his own for his $300 machine — simpsoka · 2026-10-24
- From Copilot to Claude Code: the small powerful ideas behind AI coding UIs — vinvan · 2026-10-24
- Stackchan robot runs Nous Research's Hermes Agent in YOLO mode — alexcovo_eth · 2026-10-24