多智能体团队：烧钱多，收益微
论证要点
- Vals AI 在 Vibe Code Bench 上测试 GPT-6 Sol 与 Claude Opus 5.5，单干与组队各跑中等和最高推理档，团队成本是单体的 1.8 到 5.1 倍。
- 四次团队与单体对比中仅一次有统计显著提升：GPT-6 Sol 中等推理档团队高 7.3 分；最高推理档下团队对两个模型都无实质优势。
- Anthropic 用 Opus 5.5 做的两项测试中，增加智能体数量后质量增益递减：从 10 个扩到 100 个，24 小时后分数只微升。
- Fable 5.1 在 Lean 定理证明任务上超过 10 个智能体后质量增益更强，但所有测试总分仍低于 Opus 5.5；知识库任务从 30 扩到 100 个智能体时分数反而略降。
- OpenAI 研究员 Noam Brown 在播客中确认，多智能体系统主要买速度不买质量：4 个智能体任务快一倍但成本也翻倍，16 个时模式依旧但效率略降。
- 效果高度依赖任务类型：网络研究和数学适合并行，写小说不适合；Brown 称扩展到极大规模智能体仍基本未被探索，因为成本太高。
对实践的启示
做 AI 应用时，别默认多智能体等于更强。先问任务能否并行、模型是否已跑满算力；多数场景下单体加更高推理档更划算。若要用团队，重点应放在协调成本可控的并行型任务上。
作者承认的反方与局限
作者引用 Noam Brown 指出，扩展到极大规模智能体仍基本未被探索，因为成本太高；Fable 5.1 在定理证明任务上超过 10 个智能体后质量增益更强，说明某些任务类型可能例外。
原文摘引
The results suggest that the extra cost of agent teams isn't worth it in most cases, especially when models are already running at full compute.
Four agents solved tasks twice as fast but also cost twice as much.
Throwing 10,000 agents at a novel would be just as pointless as throwing 10,000 people at it.
▸ 展开原文全文（26 段，抓取自 The Decoder）
AI agent teams waste massive tokens for barely measurable quality gains, research finds
AI agent teams deliver almost no better results than single agents, research finds.
Evals company 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.
Out of four comparisons between teams and solo agents, only one showed a statistically significant improvement: GPT-6 Sol at medium reasoning, where the team scored 7.3 points higher. At maximum reasoning, the team setup gave neither Sol nor Opus 5.5 any real advantage. The results suggest that the extra cost of agent teams isn't worth it in most cases, especially when models are already running at full compute.
Anthropic saw quality gains shrink as it added more agents in two of its own tests with Opus 5.5. Larger teams reached a given performance level faster, but going from ten to 100 agents only nudged scores up slightly after 24 hours. In separate ProgramBench tests, speed gains came with higher token usage.
| Task with Opus 5.5 | 1 agent | 10 agents | 30 agents | 100 agents |
|---|---|---|---|---|
| Knowledge base | 0.53 | 0.70 | 0.71 | 0.74 |
| Lean Theorem Proving | 0.39 | 0.66 | 0.66 | 0.68 |
Fable 5.1 showed stronger quality gains on the Lean theorem proving task above ten agents, but still scored below Opus 5.5 across all tests. On the knowledge base task, Fable's score actually dipped slightly when scaling from 30 to 100 agents.
More agents buy speed but not better output
OpenAI researcher Noam Brown confirmed in the Dwarkesh Podcast that multi-agent systems mainly buy speed, not better quality. Four agents solved tasks twice as fast but also cost twice as much. At 16 agents, the pattern held but grew slightly less efficient.
The effect depends heavily on the task. Web research and math parallelize well, but writing a novel doesn't, he said. Throwing 10,000 agents at a novel would be just as pointless as throwing 10,000 people at it. Brown acknowledged that scaling to very large numbers of agents remains largely unexplored because the costs are simply too high.
OpenAI developer Eric Provencher recently warned against using agent swarms for exactly this reason. They're most likely wasted money, he argued, because coordination between agents breaks down. He called it the coordination tax.
AI News Without the Hype – Curated by Humans
Subscribe now
Read on for the full picture.
Subscribe for hype-free coverage.
- Full access to every article on THE DECODER
- No ads
- Join the comments and community discussions
- A weekly AI news recap via mail
- 6x/year: "AI Radar" — deep dives on the AI topics that matter most
- Daily AI news, always up to date
- Our full ten-year archive
- Covered by a team with 10+ years in AI