📄 中文摘要
在过去两年中，单智能体系统不断成熟，微调技术、提示工程框架和工具使用的可靠性都有了显著提升。然而，构建单一智能体的方式正面临瓶颈。真正值得构建的系统并不是单打独斗的个体，而是协调合作的团队。例如，一个研究智能体可以委派给数据抓取工具，一个规划者可以与编码者协调，销售智能体在做出承诺前会检查库存。这些并非假设，许多公司正在实际构建这样的系统，但在尝试时却遇到了许多未被广泛讨论的问题。
无人谈论的多智能体基础设施问题
出处: The Multi-Agent Infrastructure Problem Nobody Is Talking About
发布: 2026年3月8日
📄 English Summary
The Multi-Agent Infrastructure Problem Nobody Is Talking About
Over the past two years, single-agent systems have matured significantly, with improvements in fine-tuning, the emergence of prompt engineering frameworks, and reliable tool usage. However, building solely with a single agent is hitting a wall. The systems that are truly valuable are not solo performers but orchestrated teams. For instance, a research agent can delegate tasks to a scraper, a planner can coordinate with a coder, and a sales agent checks inventory before making commitments. These are not hypothetical scenarios; companies are actively building such systems. Yet, the moment one attempts to implement these multi-agent systems, they encounter numerous challenges that have not been widely acknowledged.