Agensh paper: Microsoft's 1,024-agent leaderless swarm shows emergent division of labor
aigclink · x · 2026-09-29
The paper behind Agensh, Microsoft's decentralized multi-agent framework (companion to the earlier explainer thread). It removes the orchestrator: up to 1,024 agents receive identical task briefs and self-coordinate. 128 agents outperform a single agent's 2-hour run in 30 minutes; at scale, emergent behaviors appear—interface agreements, peer review, specialized roles, an invented integration protocol, and dedicated integrator agents. Each agent loops through check-progress, claim, execute, self-verify, and PR-merge steps, on Git + messaging + shared-wiki infrastructure. Paper link in post.
More from coding & agent
- How to keep AI-built apps from looking vibe coded: designer themes plus a design MD file for agents — Roger_M_Taylor · 2026-10-23
- freeCodeCamp tutorial: how to build production-safe agent loops after a Claude Code run burned $50,000 — Roger_M_Taylor · 2026-10-23
- Frontend dev left holding two senior backend jobs after company bet on Claude Code — ParticularFan1122 · 2026-10-23
- Sentry ships official MCP server letting AI assistants query errors, manage projects — modelcontextprotocol · 2026-10-23
- YouTube Trends MCP serves search interest and growth metrics to AI assistants — modelcontextprotocol · 2026-10-23
- Karpathy-inspired harness prompt turns Opus 5.5 into a self-verifying execution engine — Roger_M_Taylor · 2026-10-23