Amazon 论文：企业工作流中单智能体通常优于多智能体架构
做企业自动化的人容易被“多智能体协作”吸引，这篇 Amazon 的实测数据很有说服力：多智能体要多花最多 4.7 倍 token，首个版本还打不过单智能体。先用单智能体跑通、修 bug，比一上来就搭编排架构划算得多。
Amazon 发布新论文，指出在企业工作流场景中，单智能体（single agent）通常是更好的默认选择，因为多智能体架构成本高出许多，收益却有限。研究团队在 Claude Opus 4.8 上测试了 4 种智能体设计的 8 个变体，覆盖 13 条企业业务流程，任务和模拟工具保持固定，只改变架构。
结果显示，所有多智能体或技能式配置消耗的 token 是单智能体的 1.4 倍到 4.7 倍，多数情况下收益很小。首个多智能体版本的任务成功率为 58.2%，低于单智能体的 67.4%；修复后的版本虽然胜出，但 token 成本也最高。
论文给出的建议是：先从单智能体起步，逐条工作流测量成功率，先修复实现层面的 bug，再考虑是否值得为更复杂的架构付出额外成本。
New Amazon paper finds that for enterprise workflows, a single agent is usually the best default, because multi-agent setups cost far more for small gains.
Teams automating long, branching business procedures can pick a single agent, an agent with skills, an orchestrator with workers, or a swarm of specialists. Little controlled evidence exists for choosing among them.
They ran 4 agent designs, as 8 variants, on Claude Opus 4.8 across 13 enterprise procedures. Tasks and mock tools stayed fixed, so only the architecture changed.
Every multi-agent or skill setup used 1.4x to 4.7x the tokens of the single agent, mostly for small gains. A first multi-agent build lost to the single agent, 58.2% to 67.4% task success, while a repaired version won but cost the most tokens.
Start with a single agent, measure success per workflow, and fix implementation bugs before paying for a more complex architecture.
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. Key entities and technical topics under continuous tracking include 多智能体, Claude, Amazon, 智能体架构.
Source attribution: curated from rohanpaul_ai (@rohanpaul_ai).
