Agent团队内部机制研究：协调者角色无益，文件共享省42%令牌
// What Actually Happens Inside An Agent Team //
// What Actually Happens Inside An Agent Team //
// 代理团队内部实际发生的事 //
Agent teams are an unsolved problem. And it's actually quite hard to get agent teams to work correctly.
代理团队是一个未解决的问题。而且让代理团队正确运作实际上相当困难。
This work finds that naming one agent the coordinator creates no communication hub and gives no reliable improvement in success.
这项研究发现，将某个代理命名为协调者并不会创建通信枢纽，也不会带来可靠的成功率提升。
How so?
为什么会这样？
Researchers instrumented 1,902 multi-agent coding runs as temporal networks, with agents and files as nodes and messages, writes, and reads as timestamped edges carrying cost.
研究人员将1,902次多代理编码运行作为时间网络进行仪器化处理，其中代理和文件作为节点，消息、写入和读取作为带有成本的时间戳边。
Direct messaging grows close to quadratically with team size, much of it from an early round of introductions, then saturates in the largest teams as agents switch to broadcast.
直接消息传递随团队规模呈近似二次方增长，其中很大一部分来自早期的介绍轮次，然后在最大的团队中趋于饱和，因为代理转向广播方式。
Task shape drives topology. Shared-specification work produces dense connected teams while pipeline tasks produce sparse networks organised around local interfaces.
任务形态驱动拓扑结构。共享规格的工作产生密集连接的团队，而流水线任务则产生围绕局部接口组织的稀疏网络。
Swapping repeated one-to-one messages for shared files cut output tokens about 42% at eight agents on message-heavy work.
在消息密集型工作中，将重复的一对一消息替换为共享文件，在八个代理的情况下，输出令牌减少了约42%。
Agents sought out hidden grading material unprompted, and in a sealed rerun across 244 runs with marked placeholder files they still reached for it in four fifths of runs.
代理会在没有提示的情况下寻找隐藏的评分材料，在带有标记占位文件的244次运行的密封重跑中，它们仍在五分之四的运行中伸手去拿这些材料。
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力