PapersSelected

A²-Failures:多智能体系统 20 种失效模式分类法发布

Why we picked it

搞多智能体协作的都该看看这套分类法。2500 多次运行归纳出 20 种失效模式,比如浏览器被屏蔽后智能体连 shell 都没试就放弃,而且 2500 次里没有一次有人核验队友的说法。论文、代码、数据全公开,直接拿来对照自己的系统排查。

SFResearch 团队发布 A²-Failures,一套针对多智能体系统(每个智能体服务于不同的人或组织)的失效模式分类法,共归纳 20 种失效模式,基于 2,500+ 次运行的数据。论文、代码和数据集均已公开,配套博客同步发布。

Body

测试中最典型的案例是:让一组智能体回答 2026 年世界杯冠军,只有其中一个能联网。当它的浏览器工具被屏蔽后,它直接告诉队友自己没有网络访问能力,没有尝试本可成功的 shell 命令,整组最终得分为零。另一次运行中,同角色的智能体用 curl 查到西班牙夺冠,团队提交了正确答案。

数据揭示的核心问题是:2,500+ 次运行中,没有任何智能体独立核验队友的说法,导致一个智能体的错误认知直接变成团队结论。作者将这类"能力与自我认知之间的差距"列为分类法中的失效模式之一,对构建跨组织协作智能体系统的开发者有直接参考价值。

Original · Salesforce AI Research

We asked a team of AI agents who won the 2026 World Cup. Only one of them could reach the internet, but when its browser tool was blocked, it told its AI teammates it had no web access and never tried the shell command that would have worked; the run scored zero.

In another run, an agent in the same role used curl, found that Spain won, and the team filed the right answer.

That gap between what an agent can do and what it believes it can do is one of 20 failure modes in A²-Failures, our taxonomy of how agent teams break down when each agent works for a different person or organization. Across 2,500+ runs, no agent independently checked a teammate's claim, so one agent's mistaken belief became the team's conclusion.

🌎 Blog: https://t.co/ltvzkM3pzf 📜 Paper: https://t.co/n9tWjW2oCu 💻 Code: https://t.co/noa1Oa4Onk 🤗 Data: https://t.co/nX1UMtKBMs

#AIAgents #MultiAgentSystems

Industry Coordinates & Source Verification

This report is categorized under AITOP's "Papers" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows and US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include multi-agent systems, A²-Failures, agents, AI safety.

Source attribution: curated from Salesforce AI Research (@SFResearch).