🏢 Put 1,024 AI agents to work with no central coordinator, and an organization nobody designed starts assembling itself. Title: Agensh: Scaling Organizational Intelligence to 1,024 Agents URL: 📋 Overview From Microsoft Research, a self-organized multi-agent harness with no central orchestrator. Workers run concurrently and asynchronously, sharing their progress through a thin layer of shared infrastructure. 🔥 The problem it solves Most existing multi-agent harnesses use an orchestrator that decomposes tasks and assigns them out. That caps the whole system's scalability at one coordinator's capacity to manage workers and integrate what they produce. 🛠 Methodology Every worker runs the same five-step loop: gather context, claim a sub-task for itself, take action, verify against acceptance criteria, merge into the shared workspace, repeat. Three pieces of infrastructure carry it: a shared workspace on Gitea, messaging on Mattermost, and an append-only shared context holding typed entries for observations, confirmed facts, failed approaches, and active claims. The cooperation protocol lives in each worker's prompt rather than the runtime, so it layers thinly over Claude Code or Copilot. 📊 Results Tested on the five hardest ProgramBench tasks, rebuilding software from scratch with no internet access and a 6-hour budget. ・Mean test-pass rate rises from 19.31% with 1 agent to 28.78% with 128, roughly a 49% relative gain ・On pandoc it reaches 55.06% with 1,024 agents, up from 33.89% with one ・Time-to-score shifts too: 128 agents clear 30% on pandoc by the 30-minute mark, 32 agents at 60, 8 agents at 90, while a single agent never gets there within two hours 🌱 Cooperation that emerged Every worker gets an identical prompt, yet the shape of cooperation changes with scale. At 8 they agree on interfaces and split the work; at 32 approvals get withdrawn and re-reviewed; at 128 they pick reviewers by past experience and standardize a merge protocol; at 1,024 an integrator role appears, with workers polling several candidates and taking whoever answers first. #MultiAgent# #AIAgents#