Microsoft 论文：无中心管理者的编程智能体团队表现更好
Microsoft 发了篇论文，说编程智能体不用主控也能干活，128 个智能体自己认领任务、走 Git 合并，分数一路涨。
Microsoft 的论文提出 Agensh 框架，去掉主控智能体，让每个智能体自行认领子任务，在共享 Git 仓库中开发、测试、合并，并通过共享看板记录进展。在 ProgramBench 的 200 个任务中最难的 5 个上，团队规模从 1 个扩展到 128 个智能体，平均分数逐步提升。相比之下，主流多智能体编程工具都要经过 1 个主控智能体分发任务，可管理的辅助数量有限。论文所有实验使用同一模型，未报告大规模团队的成本。
Turns out coding agents don't need a boss:
New Microsoft paper finds that bigger teams of coding agents score higher and get there sooner when agents claim their own tasks without a central manager.
Even on the 5 hardest of ProgramBench's 200 tasks, scaling a manager-free team from 1 to 128 agents raised the average score at every step.
so for big jobs, add agents and let them coordinate through shared tools with no lead agent.
Popular multi-agent coding tools send all work through 1 lead agent, which can only manage so many helpers.
Agensh drops the lead agent. Each agent claims a sub-task, builds and tests it, merges it into a shared Git repo, and logs findings on a shared board.
Every run used 1 model, and the paper does not report what large teams cost.
- DAIR.AI10-04 15:38原文
