disrouter.aiAI 情报台

disrouter.aiAI 情报台

图片:The Decoder

行业新闻The Decoder

研究称 AI Agent 团队收益有限

多 Agent 编排显著增加 token 成本,但质量提升很小。

为什么重要

为开发者评估多 Agent 系统的成本收益提供了实测参考。

阅读原文the-decoder.com

AI 解读(根据原文生成)

The Decoder 报道称,评测公司 Vals AI 在 Vibe Code Bench 上测试 GPT-6 Sol 与 Claude Opus 5.5 的单 Agent 和团队设置,并比较 medium 与 maximum reasoning effort。团队成本为单 Agent 的 1.8x 到 5.1x,但四组对比中只有 GPT-6 Sol 在 medium reasoning 下出现统计显著提升。文中还引用 Anthropic 与 ProgramBench 相关测试,以及 OpenAI 研究者和开发者的观点,认为多 Agent 主要换来速度而非更好输出,并存在协调成本。

要点(AI 据原文整理,数字均已与原文核对)

  • Vals AI tested GPT-6 Sol and Claude Opus 5.5 on the "Vibe Code Bench," both solo and as teams, at two reasoning levels: medium and maximum reasoning effort.
  • The teams cost between 1.8x and 5.1x more than single agents.
  • Only one of four comparisons showed a statistically significant improvement: GPT-6 Sol at medium reasoning scored 7.3 points higher.
  • At maximum reasoning, the team setup gave neither Sol nor Opus 5.5 any real advantage.
  • With Opus 5.5, the Knowledge base task scored 0.53 for 1 agent, 0.70 for 10 agents, 0.71 for 30 agents, and 0.74 for 100 agents.
  • With Opus 5.5, Lean Theorem Proving scored 0.39 for 1 agent, 0.66 for 10 agents, 0.66 for 30 agents, and 0.68 for 100 agents.

全部来源(1)

  1. [1]The Decoder

    研究称 AI Agent 团队收益有限