# 多智能体协作研究：8个智能体时任务全失败

- 来源：Rohan Paul (@rohanpaul_ai)
- 发布时间：2026-08-23 19:52
- AIHOT 分数：41
- AIHOT 链接：https://aihot.virxact.com/items/cmt5rwnoa0p88ro73ef8rsxtd
- 原文链接：https://x.com/rohanpaul_ai/status/2091493725637857664

## AI 摘要

一项对1,902次AI编码智能体运行的研究发现，任务在2个和4个智能体时10次通过9次，但在8个智能体时全部失败。失败源于一个关于取整的规则落在两个智能体之间的决策空隙中，无人负责。研究提醒，添加智能体前应明确哪些决策会被推入协作间隙。

## 正文

Splitting a task across more agents also splits decisions that used to sit inside one agent's head.

Adding agents to a coding task moves decisions into the gaps between them.

That's where these teams break.

A study of 1,902 AI coding agent runs found one task passing 9 of 10 runs at 2 and 4 agents.

At 8 agents, one step each, it failed all 10.

A single rule about rounding fell between two of them.

In smaller teams one agent held both sides and settled it alone.

Split apart, neither owned it, and the teams raised it in every failed run without agreeing.

A prompt line saying "you are the coordinator" builds no structure in an agent team.

Before adding an agent, ask which decision you're pushing into the gap between two of them.

– arxiv. org/abs/2608.16801

Title: "When Agents Coordinate: Measuring Coordination in Multi-Agent AI Coding"