When an Agent Should Keep Researching or Start Answering
This tutorial explains how an AI Agent can decide whether more research is needed, whether to answer with qualifications, or whether to stop after hitting execution limits. Using a constructed ETF share-comparison example, it covers evidence gaps, conflicting documents, bounded research, and the separation of verified facts, assumptions, calculations, and judgments. The article also notes that the example is pedagogical rather than a live fund task, while its engineering mechanisms come from the ETF Agent design.
The article offers a practical framework for preventing research agents from either over-searching after the relevant facts are clear or withholding partial conclusions unnecessarily.
用户问:“同一只基金的两种份额,新的那个年费率更低,我是不是应该换过去?”
一个只擅长收集信息的 Agent,可以很快找到费率表、基金公告、产品介绍和几篇分析,然后写出一份资料丰富的回答。但它仍可能没有解决最重要的问题:较低的持续费用,能否在用户打算持有的时间里,抵消这次转换产生的成本?
另一种失败恰好相反。系统要求“至少查询若干来源、列出所有风险、再经过多个专家”,结果关于费率的事实已经清楚,Agent 仍不断搜索,直到没有时间组织结论。
本文关注的是两者之间的判断:怎样决定下一步还有没有必要,怎样在条件不足时给出有用回答,以及怎样把新证据真正综合成结论。以下费用、份额名称与研究过程均为教学构造,工程机制来自 ETF Agent 的设计与实现。
图示主题:从用户问题到有条件研究结论的决策链。概览按职责组织,不是某次任务的真实调用记录。 完整配图见文末原文链接。