` 一份报告出现在页面上，并不意味着一次使用已经完成。
用户还可能不知道：系统有没有读懂自己的限制，等待时能不能问别的问题，刚才补充的要求是否被采用，结论中的数字从哪里来，以及关闭页面以后去哪里继续看。
Agent 客户端要组织的，是一段可理解、可干预、可续接的协作过程。 好的体验不仅让结果出现，也让用户在过程中始终知道自己交代了什么、现在能做什么，以及下一步应该怎样判断。
下面用一个教学情境贯穿全文：
“比较我选的两只基金，重点看持仓重合和费用，先给我关键差异。我可能会离开一会儿，回来继续看。”
这里不使用真实基金数据，也不推导投资结论。ETF Agent 已有的并发对话、渐进回答、引用阅读和异步交付实践提供工程背景；本文进一步讨论交互取舍。示例文案与布局是设计表达，不代表每个细节都已在产品中实现。
整条使用路径可以概括为：提出需求 → 确认范围 → 研究与阶段阅读 → 必要时补充或停止 → 核查与取得结果 → 离开后重开 → 继续追问。 这些环节允许交错发生，用户不必等到最后一步才获得价值。
提出需求、确认范围、边等边读、调整、核查与重开六个交互时刻
用户旅程概览：这些动作可以交错发生，不是要求每次提问通过六道关卡。
一、从一个问题开始，而不是从一张配置表开始
对于这项基金比较，用户首先想表达的是“比较什么、重点看什么”。输入入口应该帮助他说清楚，而不是要求他先理解研究深度、工具组合或模型设置。
最基本的界面可以仍然是聊天框，但要让已经选中的对象和材料可见。例如，输入区附近展示两只基金的名称、已选资料和本次研究的范围；用户可以移除选错的材料，也能看出哪些个人信息将被用于这次请求。
可见的上下文比隐形的自动理解更重要。 如果系统悄悄沿用旧会话里的基金、日期或账户，用户很难发现后面的分析已经偏离了本意。
资料进入过程也需要准确反馈。文件上传完成，最多说明传输结束；文件是否能够读取、是否只有部分内容可用，是另一件事。界面不宜在上传结束时就显示“已完成分析”。失败时应保留用户的问题，并把无法使用的材料单独指出。
不过，不必为了严谨，把每次提问都变成必填表单。明确的对象与限制可以随消息一起展示；确实会改变结论、又无法合理确定的条件，再进入澄清。
三种空间，围绕同一项任务协作
对话、任务卡和报告，分别回答不同问题。对话保留用户问了什么、改了什么；任务卡说明现在到哪一步、还能怎样干预；报告负责解释发现，并为核查依据提供稳定位置。它们不必变成三个独立页面，更不应各维护一份需要手工同步的研究结果。
可以把任务卡放在对话上方，只写“基金 A 与基金 B · 持仓比较待补齐”，旁边提供“阅读已有结果”“补充要求”和“停止”。报告仍是主要阅读去处。用户转去问一个无关的术语时，任务卡让原研究仍然找得到，却不遮住新回答。这是一种信息布局建议，不代表新增页面已经实现。
在手机上，优先露出任务名称、影响判断的缺口和一个最有用的动作，次要操作再收起。详细过程可以按需展开；来源详情可以临时替换阅读区，不必勉强并排。关闭详情后回到同一句结论。窄屏需要保留的是任务关系，不是桌面的列数。
同一研究的任务卡、结果正文和对话输入怎样协作
教学布局，不是产品截图。任务卡帮助干预，正文帮助判断，对话帮助表达新的意图。
二、需要问用户时，只问会改变研究的条件
假设两只基金的资料披露日期不同。系统可能需要确认：用户想按最近可取得的资料比较，还是要求严格对齐某个时间点？这会影响结论能否成立，值得问清楚。
相反，“你喜欢详细还是简洁”“是否需要表格”“要不要风险提示”并不总值得逐个阻塞研究。能够采用合理默认的展示偏好，可以先执行，再允许用户调整。
好的澄清应同时给出问题的原因和可选择的路径。例如：
“两份持仓资料的日期不同。可以先比较各自最近披露的数据，并标明日期差异；也可以等你指定同一个比较时点。”
这比单独问“请选择时间范围”更容易回答，因为用户知道自己的选择会改变什么。
用户回复以后，也应能看见采用后的研究范围。没有必要让他重新读完整聊天记录，才能确认系统是否仍然在比较正确的对象。
对于依赖用户回复的任务，可以显示“等待你确认比较时点”；对于能够继续完成独立部分的任务，则可先交付不依赖该条件的内容。是否能继续，取决于任务本身，不应由一个统一的“等待中”画面决定。
三、等待时，让用户理解进展，而不是观看运行日志
研究开始以后，最容易做的是展示一个不断变化的状态区：正在搜索、调用工具、分析中、生成中。
它可能证明系统还在活动，却未必帮助用户判断是否值得继续等待。
对这个基金比较，真正有用的进展更接近：已经取得哪些比较材料、还有哪一项关键信息缺失，以及是否已经能回答一部分问题。比如“费用资料已取得，持仓比较还缺另一只基金同类口径的披露”，比连续几条工具调用记录更贴近用户的任务。
这里要注意措辞的来源。系统只能在确有相应事实时说“已取得”；一段计划或自述不能自动升级成核验完成。无法可靠估算剩余时间时，可以说明正在等待什么，不必制造一个不断走向 99% 的进度条。
页面可以把信息分成轻重两层：默认展示简短进展、关键阻塞和已有成果；用户主动展开时，再查看更细的过程。模型切换、请求重试和工具参数没有必要成为普通读者的必经内容。
更早出现有用回答，比更早出现一段“我将开始分析”更值得追求。 ETF Agent 的渐进回答实践也把“首次有用回答”与“首次输出文字”分开：先回答已有依据的重要子问题，后续围绕同一篇回答补充或修正，而不是每一步都再写一份小报告。[1]
于是，等待可以逐渐变成阅读。费用部分已经足以说明差异时，用户可以先看它；尚未取得的持仓材料保持明确缺口，不必等所有工作结束才展示任何成果。
第一份结果出现后，操作入口也应该改变
还没有成果时，“查看进展”可能有用；已有依据的费用比较可以阅读后，“先看费用差异”就比继续展示转圈更有意义。关键日期仍待确认时，可以把问题放在受影响的部分旁边，而不是用一张等待画面替换整份报告。
这形成了一段很小、却重要的交互推进：等待有原因，阶段成果有入口，未解决条件有回复位置。不要为了让任务卡显得活跃而加一个百分比。没有可靠进度估算时，诚实的缺口比虚构的完成比例更能帮助用户采取行动。
四、研究还在做，用户仍然应该能够说话
看到某个费用术语，用户可能顺手问一句：“这个费用和申购费有什么区别？”
这不一定是在修改原比较。要求用户先停止研究才能提问，会把协作过程重新变成单线程等待；把所有新输入都塞进原研究，又可能改变原来的任务。
因此，界面至少应该帮助用户分清四种动作：提出一个独立问题，补充原研究，回答某个待确认问题，以及停止指定任务。
不一定需要四个永久占位的输入框。一个普通聊天入口，加上研究卡片里的“补充要求”和待确认问题旁的回复入口，通常就能把大部分意图表达清楚。进入补充模式时，输入框应显示明确的目标，例如“补充到：基金 A 与基金 B 的比较”，同时允许退出这个模式。
用户说“把持有一年时的费用也放进去”，系统可以回显这条补充，并说明当前处于什么阶段。已收到不等于已经用上，用上也不等于报告已经完成更新；需要等待结果才能确认的事，不应提前写成成功。
ETF Agent 的既有并发实践采用了相近的分工：独立短问单独回答，明确的补充交给原研究，停止操作指向具体任务。用户能继续说话，并不意味着系统承诺无限并发；容量不足时，应明确说明新任务没有被接纳，不能悄悄把它并入另一项研究。[2]
停止按钮也无需消失。它应该属于对应任务，并清楚说明影响范围。收起任务面板、切换会话和关闭网页，则应当与主动停止区分开来。真正允许离页继续执行的产品，还需要把这一点明确告诉用户；不支持时就不能暗示支持。
独立提问、补充要求、回答澄清与停止任务具有不同目标
交互分流示意：目标明确时直接携带目标，不让用户每次重新解释自己正在修改什么。
五、结果开始出现时，先让人读懂，再考虑动画
阶段回答进入页面以后，用户关心的已经不只是“还在工作吗”，而是“这些内容对我的问题意味着什么”。
对这个教学案例，可以先呈现已经有依据的关键差异，再说明比较依据和仍缺的材料。结论需要保持条件完整：不能把“当前取得的资料显示……”压缩成无条件的最终判断。
正文的组织也不应完全照抄研究执行顺序。用户通常不需要按“第一次搜索、第二次阅读、第三次补充”理解两只基金。他需要的是费用怎样比较、持仓重合能说明什么，以及哪些部分暂时不能判断。[1]
布局上，可以让回答保持一条稳定的阅读主线。过程区域按需展开，来源在相关结论附近可进入，下载和后续操作放在成果旁边。窄屏上不必复制桌面的多个并排面板；先保证正文、任务状态和必要动作能顺序读到。
流式更新尤其需要尊重阅读位置。用户已向上查看前面的内容时，后续文字到达不应反复把页面拉到底部。一个“有新内容”的提示，比强制抢走滚动位置更符合正在阅读的任务。
修订也不必每次都弹窗。真正影响判断的变化，可以用简短的更新说明指出，例如“持仓比较已补齐，原先未确认的部分现在有了依据”。这类说明必须来自实际变化，不能为了制造活跃感而固定重复。
六、想核查一句话时，不要先把用户赶出报告
假设用户读到“这两只基金的某类风险暴露较为接近”。他可能只想确认：这句话依据哪些材料，有没有忽略披露日期，究竟是记录中的事实还是进一步的判断。
如果点击引用立即打开一张长网页，读者还要重新找段落、找表格，再返回报告定位刚才那句话。链接虽然存在，核查路径却没有真正完成。
更合适的入口是先在当前阅读环境里展示相关材料。原文片段、来源日期或版本、读取位置和必要限制放在一起；“打开原来源”保留为一个清楚的独立动作。当前网页可能已经更新，因此原来源页面与当时用于研究的材料不应被混称为同一份内容。[3]
不同结论还需要不同的核查方式。
事实陈述适合查看原始材料和上下文。计算结果需要看实际使用的输入、单位、口径和计算关系，而不仅是一个来源标题。分析判断则需要说明依据、假设和仍然存在的不确定性；它不能被包装成原文已经直接说过的话。
没有留存原始片段时，应当直说。不能现场生成一段貌似原文的文字填上空白；重新获取材料或额外发起分析，也应该是另一个可辨认的操作，而不是“展开详情”的隐藏副作用。
同样，一条引用失效不应自动抹掉整篇仍可合法阅读的正文，但资料不可用与访问权限撤回也不能一概而论。界面要显示正确的缺口，不得继续从旧缓存展示当前不允许查看的内容。[3]
这些入口还必须真正可操作。短详情就地展开，长材料放在侧栏或对话框，都可以成立；不能只依赖鼠标悬停。键盘用户应能打开、阅读、关闭，并回到原来的操作位置。W3C 的折叠与模态对话框模式提供了不同的交互约定，具体组件应选择一致的模式实现。[4][5]
核查的目标是让用户有条件判断这句话，不能因为界面展示了来源，就替用户宣布研究已经正确。
事实、计算和分析判断分别需要哪些可核查材料
核查路径示意：展开原有材料与重新研究是不同动作；来源可读不等于结论已经正确。
七、交付一份结果，也要让用户知道怎样取得它
有些用户在聊天页面读完就够了，有些用户需要把报告带走，或者保留一个以后能够重新打开的版本。
这时，复制正文、下载文件和保存结果，不应该被混成一个含义不明的“完成”按钮。
复制意味着取得当前可复制的文字；下载需要有实际可读取的文件；保存或收藏则应清楚说明保存的是什么。对于仍在更新的报告，用户尤其需要知道自己取得的是当前哪一版，之后的新修订是否会改变已经保存的内容。
产品支持附件时，应在结果附近展示文件名、格式和可用状态。正文已经能读，但某个附件生成失败，可以单独说明附件缺失；没有必要把整次研究判为毫无成果。反过来，一个文件保存成功，也不能证明用户已经能够打开它。
这些功能不是所有客户端都必须第一天具备。只有聊天阅读需求时，可靠的历史入口可能已经足够；确实存在交付与流转需求，再增加版本明确的下载、收藏或分享能力。分享还需要独立检查访问范围，不能把“我看得到”理解成“任何拿到链接的人都看得到”。
从阅读阶段结果、补充持有期限到取得指定版本的交接路径
一次交接示意：收到补充要求不等于报告已改；只有真实修订才提示变化，只有可读取的文件才提供下载。后续更新不应悄悄替换用户已取得的版本。
八、用户离开之后，回来应接上原来的事
用户可能在持仓比较尚未完成时关闭页面。一个承诺后台研究的产品，应让他知道任务是否继续、以后从哪里找到结果，以及需要他回复时会怎样提醒。
重新进入时，最重要的不是重放一遍打字动画，而是迅速恢复理解：这是什么任务，最后完成到哪一步，有哪些结果已经能读，还缺什么，是否需要我做决定。
页面应该读取原任务的已保存结果，而不是悄悄把原问题再发一次。否则，用户以为自己是在打开历史，系统却开始了另一项研究，内容和费用都可能改变。ETF Agent 的异步交付实践明确区分了离页、重开、通知与研究执行，历史重读不以重新生成报告为前提。[6]
通知本身也不应该变成另一份报告。它只需准确说明发生了什么，并把用户带回对应任务和结果；通知被发送，不代表用户已经读过正文。
当用户约定时间时，语言差异要反映到交互里。“十分钟后提醒我”与“十分钟内给我结果”不是同一件事。提前完成是否通知、到点仍未完成时怎样说明，都应与原约定一致。已有阶段结果就指向阶段结果，还没有结果就说明当前状态，不能把“接近完成”当成交付。[6]
最后，重新打开仍然是一项当前的访问请求。换了账户、权限变化或材料撤回以后，旧链接不应自动恢复原来的阅读资格。对用户可以只解释“当前无法访问这份材料”；不必在界面展开权限系统的内部实现。
让用户认得出自己回到了哪里
重开时，应该能看见原问题、当前可读结果、相较先前真正影响判断的变化，以及仍等待回复的事项。一个日期或“已更新”标签，不足以解释究竟是持仓部分补齐了、结论改变了，还是只有附件变得可用。知道变化时说明实际变化；不知道时展示已保存版本，不编造一段更新摘要。
通知也使用同一任务名称，并进入对应结果或待回复问题。用户有多项研究时，“你的研究有新进展”往往不够明确。关闭通知不应停止任务，阅读报告也不应悄悄把尚未回答的澄清标成已解决。
重新阅读、请求继续和重新研究是不同动作
重开路径示意：先恢复原任务的理解，再由用户选择是否继续或重新研究，不自动重新执行。
九、出现问题时，告诉用户影响和下一步
研究型产品无法把所有失败都消除，但可以避免把不同问题画成同一个红色横幅。
网络断开时，用户需要知道任务状态暂时无法更新，而不是被告知研究一定失败。某份资料无法取得时，需要知道哪项比较受到影响，而不是收到一条无法理解的工具错误。任务停止时，则需要知道还有哪些结果已经保留下来。
好的错误提示尽量回答三个问题：发生了什么、现有成果受到什么影响、接下来可以做什么。没有证据的恢复承诺，不应写进按钮文案。
尤其要谨慎使用“重试”。重新读取页面、再次提交未被接纳的问题、继续原任务、另开一次研究，是不同动作。按钮应该尽可能告诉用户将发生哪一种；不能把重复执行隐藏在一个看似无害的刷新动作后面。
某些动作可能影响外部数据或产生费用时，还要考虑前一次是否已经发生。客户端不能因为没有看到回执，就认定操作从未执行。这类不确定性需要明确处理，不能用无限自动重试掩盖。
局部失败也不应扩大成整体失败。已有可靠依据的费用比较可以继续阅读，缺失的持仓部分单独标注；但不能为了保住“完成”状态，把原本要求的持仓分析悄悄从任务里删掉。
连接中断、资料缺失和任务停止需要不同的恢复入口
失败交互示意：先说明对当前成果的影响，再给出实际支持的下一步；不可用按钮不应许诺恢复。
十、继续追问时，保留讨论对象，不偷偷改变依据
用户回来读完报告，可能问：“如果只看持有一年，刚才的结论会怎样？”
这句话有明确的讨论对象：刚才那份比较及其条件。客户端可以让用户从某段结论、某个表格或整个结果发起追问，并显示引用的是哪项内容。只把一句脱离上下文的“为什么”送出去，会让系统重新猜测。
接下来，还要分清三种追问：解释旧结论、沿用原资料改变假设并重新计算，或按新资料重新研究。解释旧结论应保留原报告的时点和依据；把持有期限改为一年，需要明确新假设及其计算结果，不能让新回答冒充原结论；重新取得资料，则要说明新的资料时点和比较条件。只有这些区别确实影响结果时，才值得显式确认，不必让所有追问都经过一张版本选择表。
反馈也应指向具体对象。用户认为某个数字有问题，和他觉得报告太长，不是同一类反馈；修改后的回答也不应继续继承针对旧版本的评价。
更不能因为用户点了赞，就默认把报告中的事实写入长期记忆，或者向他声称系统已经学会了某种投资偏好。反馈、纠错和长期保存是不同操作，需要各自可理解的边界。
至此，一项研究才真正从单次生成变成可以持续使用的成果：能提出问题，能在过程中纠偏，能核查答案，能离开再回来，也能在保留上下文的情况下继续讨论。
十一、先做一条完整路径，再决定增加多少界面
覆盖完整流程，不意味着每个产品都需要同时拥有任务中心、通知中心、来源侧栏、报告库和复杂的控制台。
对于简单问答，一个输入框、一份可读回答、必要的来源和明确的失败状态就可能足够。任务开始跨越等待与离页，才需要稳定的任务入口和历史续接；用户确实会边研究边聊天，再加入明确的并发交互；结果需要流转，再增加文件和版本化交付。
决定这些增量的，应当是用户任务中的真实断点，而不是希望把后端所有能力都展示出来。
验收也可以沿着同一个教学任务进行：用户能否确认比较对象，是否知道还缺什么，补充要求有没有指向正确研究，能否在结论附近核查依据，离开后能否回到同一份成果。页面没有报错只是起点，用户能独立完成这些动作才更接近交互目标。
回到最初的问题：用户把两只基金交给 Agent 比较，期间问了一个术语，补充了持有期限，离开了一会儿，回来读报告并核查来源。他不需要知道背后调用了多少次模型、使用了哪些执行组件；他需要的是每一步都知道发生了什么，并保有继续、调整或停止的选择。
让复杂研究变得可协作，而不把系统复杂度转嫁给使用者，这才是 Agent 客户端全景应该呈现的主线。
参考与延伸阅读
[1] ETF Agent，Agent 的渐进式回答：怎样保持研究过程连贯。
[2] ETF Agent，异步后台研究与前台并发交互实践。
[3] ETF Agent，Agent 引用系统设计：从来源签发、版本绑定到独立复核。
[4] W3C WAI-ARIA APG，Disclosure Pattern。
[5] W3C WAI-ARIA APG，Dialog (Modal) Pattern。
[6] ETF Agent，研究还没结束，用户已经离开了：Agent 怎样按时交付真实成果？。
机制深挖见配套文章《Agent 客户端不只是展示 SSE：一次研究怎样被正确呈现？》。取材日期：2026 年 11 月 2 日。
原文与交流
阅读原文 · ETF Agent 工程文集（欢迎 Star）。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
`The fact that a report appears on the page does not mean that one use has been completed.
Users may also wonder whether the system can read its own limitations, whether it can ask other questions when waiting, whether the additional requirements have been adopted, where the figures in the conclusions come from, and where to continue after the page is closed.
The Agent client wants to be organized as an understandable, interventionable and continuous collaborative process. Good experience not only brings results, but also keeps users informed about what they say, what they can do now, and what next step should judge.
The following is a full text of the teaching situation:
“By comparing the two funds of my choice, the focus is on confluence and cost, giving me a key difference. I might leave for a while and come back and keep looking."
Real fund data are not used here and investment conclusions are not derived. The existing ETF Agent dialogue, step-by-step answers, references to reading and step-by-step delivery practices provide the engineering context; interactive trade-offs are discussed further in this paper. The text and layout of the example is a design expression and does not mean that each detail has been achieved in the product.
The entire path can be summed up as: demand confirmation range research and stage reading supplement or stop verification and results if necessary re-opening after leaving. These links allow for staggering and the user does not have to wait until the last step to obtain value.
Presentation of needs, confirmation of scope, side reading, adjustment, verification and reopening of six interactive moments
Overview of the user ' s journey: These actions can be staggered and not required to pass six crossing points per question.
I. Starting with a problem rather than with a configuration table
For this Fund comparison, the user wishes first to express “what to compare and what to focus”. The entry point should help him clarify, not require him to understand the depth of the research, the mix of tools or the model setting.
The most basic interface can remain a chat box, provided that the selected objects and materials are visible. For example, the name of the two funds, the selected information and the scope of the study are displayed near the input area; users can remove the wrong material and see which personal information will be used in the request.
The visible context is more important than an automatic invisibility. If the system quietly follows the funds, dates or accounts of the old session, it is difficult for users to find that the analysis that follows is out of context.
Accurate feedback is also required for the information entry process. The upload of the file is complete, indicating the end of the transfer at the most; whether the file is readable and only partially available is another matter. The interface should not display " completed analysis" at the end of the upload. The problem of the user should be preserved in the event of failure and the unusable material singled out.
However, it was not necessary to make every question a mandatory form for the sake of rigour. Clear targets and restrictions can be displayed with the news; indeed, they change conclusions without reasonable certainty before they are clarified.
Three spaces around the same mission
Dialogue, mission cards and reports, respectively, answer different questions. The dialogue retained the question of what had been asked and what had been changed; the mission card described where to go and how to intervene; and the report explained the findings and provided a stable location for the verification basis. They need not turn into three separate pages, let alone maintain a single study that requires manual synchronization.
The task card could be placed above the dialogue, with only “Fund A and Fund B. Holdings to be completed”, with “reading the results” “additional requirements” and “stop”. The report remains the main reading site. When the user turns to a irrelevant term, the taskcard allows the original study to be found without covering the new answer. This is an information layout proposal and does not mean that the new page has been implemented.
On the mobile phone, priority is given to displaying the name of the mission, gaps in impact judgement and one of the most useful actions, and secondary operations are re-opened. The detailed process can be carried out on an as-needed basis; the source details can temporarily replace the reading area, without having to wait in parallel. The details were closed and returned to the same conclusion. The narrow screen requires the task relationship, not the number of columns on the desktop.
How to collaborate on mission cards, outcome body and dialogue input from the same study
Teaching layouts, not product screens. The cards help to intervene, the text helps to judge, and dialogue helps to express new intentions.
II. Questions for users only to change the conditions of the study
It is assumed that the disclosure dates for the two funds are different. The system may need to confirm whether users want to compare the information recently available or require strict alignment at a given point in time. This affects the validity of the conclusions and is worth asking.
On the contrary, “Do you like detail or simplicity” “Do you want a form” “Do you want a risk alert” is not always worth studying one by one. Reasonable default display preferences can be used to implement them before allowing users to adjust.
A good clarification should be given both the reasons for the problem and the alternative path. For example:
“The dates for the two holds vary. It is possible to compare the data recently disclosed by each of them and to indicate differences in dates; it is also possible to wait until you have specified the same point of comparison.”
This is easier to answer than asking “select the time frame” alone, as users know what their choice will change.
Upon user response, the scope of the post-implementation study should also be seen. There is no need for him to reread the full chat record to confirm whether the system is still the correct object.
For missions that rely on user responses, “waiting your confirmation of the date of comparison” can be shown; for tasks that can continue to be performed in a stand-alone segment, content that does not depend on the condition can be delivered first. Continuation depends on the task itself and should not be determined by a single “waiting” image.
iii. When waiting, to allow users to understand progress rather than watch the running log
The easiest thing to do when the study begins is to show a changing status zone: searching, calling tools, analysing and generating.
It may prove that the system is still active but does not necessarily help users to judge whether it is worth continuing to wait.
Compared to this fund, real useful progress is closer: what comparative information is available, what key information is missing, and whether some of the questions can be answered. For example, “cost information is available and there is a lack of disclosure of the same calibre of the other fund for the holding of the warehouse”, which is closer to the user's task than successive tool call records.
Here, attention is drawn to the source of the wording. The system can only say “achieved”; a plan or self-statement cannot automatically be upgraded to a verification. When it is not possible to reliably estimate the remaining time, it is possible to explain what is waiting for, without creating a progress line that is moving towards 99 per cent.
The page can divide the information into two layers: the default display of short progress, key obstructions and results already achieved; and the user takes the initiative to view a more detailed process. Model switching, requests for retry and tool parameters need not be part of the ordinary reader.
An earlier, useful answer is more worthwhile than an earlier “I will start to analyse”. ETF Agent's progressive response practice also separates the “first useful answer” from the “first output text”: first to answer important sub-questions on which there is a basis, then to supplement or amend the same answer, rather than to write a sub-report at every step. [1]
So, waiting can gradually become reading. When the cost component is sufficient to account for the difference, the user can look at it first; the unobserved warehouse material maintains a clear gap and does not have to wait until all work is completed before showing any results.
After the first result, the access should change.
In the absence of results, it may be useful to “see progress” and “see cost differences first” will be more meaningful when the costs are more readily readable than continue to show circles. Key dates could be confirmed by placing the issue next to the affected part, rather than replacing the entire report with a one-shot waiting image.
This led to a small but important interaction: waiting for reasons, entry points for phase results and response points for unresolved conditions. Do not add a percentage in order to make the mission card active. In the absence of reliable estimates of progress, an honest gap is more conducive to action by users than an imaginary completion rate.
Research is still under way and users should still be able to speak.
Seeing a certain term of cost, the user may ask the following question: “What is the difference between this cost and the requisitioning fee?”
This is not necessarily a modification of the original comparison. Users are required to stop research before asking questions, and the collaborative process will be re-engineered into a single-way waiting process; all new inputs will be inserted into the original study, with the possibility of changing the original mission.
Therefore, the interface should help users to distinguish between at least four actions: a stand-alone question, a supplement to the original study, an answer to a question to be confirmed, and the discontinuation of the assignment.
Four permanently occupied input boxes are not necessarily required. A regular chat portal, together with the “additional requirements” in the study card and the response point next to the question to be confirmed, usually conveys most of the intent. When entering the supplementary mode, the input box should show clear objectives, such as " Supplement to: Fund A versus Fund B" and allow exit from the model.
The user states that “the cost of holding a year is also included”, and the system can resonate with this addition and indicate the current stage. The receipt does not mean that it has been used or that the report has been updated; what needs to be confirmed is not written in advance.
The existing parallel practice of ETF Agent uses a similar division of labour: independent questions are answered separately, explicit additions are given to the original study and no operation is directed to specific tasks. The fact that users can continue to speak does not mean that the system is committed to an unlimited parallel; when the capacity is insufficient, it should be made clear that the new task is not accepted and cannot be quietly incorporated into another study. [2]
Stop buttons do not have to disappear. It should fall within the context of the corresponding mandate and clearly define the scope of its impact. Close task panels, switch sessions and close web pages should be distinguished from active stoppages. A product that is truly allowed to continue on the page also needs to be clearly shared with users; support cannot be implied without support.
Independent questions, additional requests, answers to clarifications and terminations have different objectives
Interactive diversion: Carrying the target directly when the target is clearly defined, so that users are not allowed to repeat what they are modifying.
V. When results begin to appear, read them and then consider animation
After this stage of replying to the page, the user is already concerned not only with “are they still working”, but with “what these content mean to my questions”.
This teaching case can be followed by a presentation of the key differences on which the evidence is based, a description of the basis for comparison and the material still lacking. The conclusions need to remain complete: the “information currently available” cannot be reduced to an unconditional final judgement.
Nor should the organization fully follow the order of implementation of the study. Users usually do not need to understand the two funds in terms of “first search, second reading, third replenishment”. What he needed was a comparison of costs, what a silo overlap would explain and what elements could not be judged for the time being. [1]
The layout allows the answer to maintain a stable reading line. Process areas are developed as needed, with sources accessible near the conclusions and downloads and follow-up actions next to the results. There is no need to copy multiple side panels on the narrow screen; first, you can ensure that the body, the task state and the necessary actions are read in sequence.
In particular, it is necessary to respect the reading position. When the user has looked up the front page, the following text should not be reached repeatedly. A hint of “new content” is more in line with the task being read than to force a rolling position.
It's not always a window. Changes in real impact judgement can be indicated in a short update, for example, that “the warehouse comparison has been completed and the previously unrecognized portion is now justified”. Such statements must come from real changes and cannot be fixed to repeat in order to create a sense of dynamism.
If you want to verify a sentence, do not expel the user from the report.
It is assumed that the users read “the exposure of some of the categories of the two funds is closer”. He might simply wish to confirm what material the sentence was based on, whether the date of disclosure had been ignored, whether the facts in the record or further judgement had been made.
If you click a reference to open a long web page immediately, the reader will also have to search for paragraphs, tables, and return to the reporting position. Although links exist, the verification path is not really complete.
A more appropriate entry point is to display relevant materials in the current reading environment. Original text clips, source date or version, reading location and necessary limitations are combined; “open source” is retained as a clear independent action. As the current web page may have been updated, the original source page should not be referred to as the same content as the research material used at the time. [3]
Different findings also require different verification modalities.
The statement of facts is appropriate for the original materials and context. The result of the calculation needs to look at the input, unit, calibration and computational relationship actually used, not just a source title. The analytical judgement requires a statement of the basis, assumptions and remaining uncertainty; it cannot be packaged in the words that are already spoken directly.
In the absence of the original clip, it should be stated. A text-filling blank cannot be generated in the field; re-acquisition of material or additional initiation analysis should also be another identifiable operation rather than a hidden side effect of “examining details”.
Similarly, the ineffectiveness of a reference should not automatically erase the body of the entire text, which could still be legally read, but the information could not be used in general terms with the withdrawal of access to permission. To display the correct gap, the interface will not continue to display what is not currently allowed to view from the old cache. [3]
These entrances must also be truly operational. Short details are then carried out in place, and long materials can be set up in sidebars or dialogue boxes; they cannot rely solely on mouse suspension. Keyboard users should be able to open, read, close and return to their original operational position. The W3C folding and model dialogue box model provides for different interactive engagements, and specific components should choose a consistent mode of realization. [4] [5]
The objective of verification is to allow users to judge the phrase and not to announce that the research is correct for users because the interface shows the source.
Facts, calculations and analyses determine which verifiable material is required
The path map for verification: it is a different move to start with the original material and to re-examine it; the source is not readable as if the conclusion is correct.
VII. Delivering a result and letting users know how to get it
Some users read enough on the chat page and some need to take the report away or keep a version that can be reopened later.
At this point, copying the body, downloading the file and saving the results should not be mixed into an unknown “complete” button.
Reproduction means obtaining current replicable text; downloads require actually readable documents; saving or collection should clearly indicate what is saved. For reports that are still being updated, it is particularly important for users to know which version they have acquired and whether the subsequent new revisions will change the content already stored.
When the product supports the annex, the file name, format and availability should be displayed near the result. The text could already be read, but the fact that an annex had failed to produce could be described separately as missing; there was no need to treat the entire study as fruitless. Conversely, the success of a document is not proof that the user has been able to open it.
Not all clients must have these features on the first day. Only when there is a chat and reading demand may there be enough reliable access to history; there is a demand for delivery and flow, with a clear download, collection or sharing capability. Sharing also requires an independent examination of the scope of the visit and cannot be understood as “i can see” to mean “i can see anyone who gets a link”.
Path from reading, supplementary holding period to the transfer path to the specified version
A single handover signal: The receipt of a supplementary request does not amount to a change in the report; only a genuine revision suggests a change and only readable documents are downloaded. Subsequent updates should not quietly replace versions already available to users.
When the user leaves, he should be back with the original.
The user may close the page when the warehouse comparison is not completed. A product committed to back-office research should provide him with information about the continuation of the mission, where to find the results later, and how to alert him when he is required to respond.
When re-entering, the most important thing is not to replay an animation, but to quickly regain understanding: What is this task, what steps have been taken to finalize, what results have been read and what is missing, and whether a decision is needed from me.
The page should read the saved result of the original task and not send the original question again quietly. Otherwise, users thought they were opening their history and the system started another study, which could change both content and cost. ETF Agent ' s step-by-step delivery practice clearly distinguishes between off-page, re-opening, notification and research execution, and re-reading history does not presuppose the re-generation of the report. [6]
Nor should the notification itself become a separate report. It simply needs to state exactly what happened and to bring back the user to the corresponding task and results; the notice is sent and does not represent that the user has read the text.
Language differences are reflected in the interaction when the user agrees on the time. "Tell me in ten minutes" is not the same thing as "Give me results in ten minutes." It should be consistent with the original agreement to indicate whether notice of early completion or, if not, the point at which completion is still pending. The results already at the stage point to the results at the stage, and the current status is illustrated without results, and “near completion” cannot be considered delivery. [6]
Finally, reopening remains a current request for a visit. The old link should not automatically be restored to the original reading qualification after the account was changed, permission changes or material was withdrawn. For users, it is possible to explain only that “the material cannot be accessed at this time”; it does not need to be internalized in the permission system in the interface.
Let the user know where he's gone.
When reopening, it should be possible to see the original problem, the current readable results, changes in the actual impact judgement compared to the previous one, and matters still awaiting response. A date or “updated” label is not sufficient to explain whether the hold-up has been completed, the conclusions changed, or only the annexes became available. Displays the actual change when you know it; displays the saved version when you do not know it, without creating an update of summary.
The notification also uses the same task name and enters the corresponding result or questions to be answered. When users have multiple studies, it is often not clear that there are new developments in your research. The notice of closure should not be discontinued, nor should the report be read quietly to mark the outstanding clarifications as settled.
Rereading, requesting and re-examining are different actions.
Reopen the path icon: First restore the understanding of the original task and then the user chooses whether to continue or re-examine, without automatic re-execution.
ix. When problems arise, inform users of influence and next step
Research products could not eliminate all failures, but could avoid drawing different issues into the same red banner.
When the network is disconnected, the user needs to know that the task is temporarily unnewable, rather than being told that the research will fail. When a particular information is not available, it is necessary to know which is more affected than to receive an incomprehensible tool error. At the end of the mandate, it was necessary to know what other results had been retained.
A good error reminder answers as much as possible three questions: what happened, how the results are affected, what can be done next. A promise of reinstatement without evidence should not be entered in the button file.
In particular, “retry” should be used with caution. Re-reading the page, re-submitting unaccepted questions, continuing the original task and opening a separate study are different actions. The button should, to the extent possible, tell the user what will happen; it cannot hide repeated execution behind a seemingly harmless refreshing action.
When certain actions may affect external data or incur costs, it is also necessary to consider whether the previous event has taken place. The client cannot conclude that the operation was never performed without a reply. Such uncertainties need to be explicitly addressed and cannot be covered by an unlimited automatic retry.
Nor should partial failure be extended to total failure. The cost comparison, where there is a reliable basis, can continue to be read, with the missing hold-up portion being singled out; however, the original hold-up analysis cannot be removed quietly from the mandate in order to preserve “complete” status.
Interrupted connections, missing information and mission stopovers require different recovery entrances
An interactive indication of failure: a description of the impact on the current outcome followed by next step, which actually supports it; no button should be used to promise recovery.
X. CONTINUING THE QUESTION, reserving the subject of the discussion and not changing the basis
The user came back to read the report, and he might ask, "What would happen if it was only for a year?"
This sentence is the subject of a clear discussion: the comparison and its terms. The client can allow the user to initiate a query from a paragraph, a table or the whole result, and show which is quoted. Sending out only one sentence of "why" away from context would give the system a new guess.
There are then three kinds of queries to be distinguished: an explanation of the old conclusions, a recalculation of assumptions using the original information, or a re-examination of the new information. The interpretation of the old conclusions should preserve the timing and basis of the original report; change the holding period to one year, which would require the clarification of new assumptions and their calculations, and not allow new answers to pretend as conclusions; and re-access to information, which would indicate the timing of the new information and the conditions for comparison. It is only when these differences do affect the outcome that they are worthy of explicit confirmation and it is not necessary for all queries to go through a version selection form.
Feedback should also target specific audiences. Users considered a number to be problematic and he felt that the report was too long and not the same type of feedback; revised responses should not continue to inherit evaluations for the old version.
Even less, the fact that the report was written into long-term memory by default, or the fact that the system had learned some investment preference, should not be used as a reason for the user ' s praise. Feedback, error correction and long-term preservation are different operations and require the boundaries that each understands.
It is thus that a study is truly moving from a single generation to a sustainable one: to ask questions, to correct them in the process, to verify the answers, to leave and come back, and to continue the discussion with context.
Eleven, make a full path and decide how much to add. Noodles.
Covering the full process does not mean that each product needs to have both a mission centre, a notification centre, a source sidebar, a reporting library and a complex control desk.
For simple questions and answers, an input box, a readable answer, the necessary sources and a clear state of failure may be sufficient. Tasks begin to cross wait and leave pages, requiring stable mission entry and history; users do talk while they research, adding clear and interactive; and results need to flow, adding document and version delivery.
These increases should be determined by the true breakpoint of the user ' s mission, rather than by the desire to demonstrate all the capabilities of the back end.
acceptance check can also follow the same teaching mission: whether the user can identify the comparator, whether it knows what is missing, whether the additional requirement is to point to a proper study, whether it can verify the basis of the conclusions in the vicinity and whether it can return to the same outcome after leaving. There are no errors on the page but the point of departure, so that the user can do these actions independently to get closer to the interactive target.
Returning to the original question: The user gave the two funds to Agent for comparison, during which a term was asked, the holding period was added, and a few days were left to read the report and verify the source. He did not need to know how many models were transferred behind them and what implementation components were used; he needed to know what had happened at every step and to retain the option of continuing, adjusting or stopping.
Making complex research collaborative rather than transferring the complexity of the system to users is the main line that Agent client landscapes should be presented.
References and extended reading
[1] ETF Agent, Agent ' s progressive answer: How to keep the research process consistent.
[2] ETF Agent, Cross-step Backstage Research and Co-execution.
[3] ETF Agent, Agent citation system design: from source issue, binding to independent review.
[4] W3C WAI-ARIA APG，Disclosure Pattern。
[5] W3C WAI-ARIA APG，Dialog (Modal) Pattern。
[6] ETF Agent, the study is not over and the user has left: How does Agent deliver real results on time?
The mechanism found the companion article "Agent client is more than just displaying SSE: How is a study presented correctly? " Date of material extraction: 2 November 2026.
Original language and communication
Read the original ETF Agent Engineering Collection (welcome Star).
Auto-fetched and translated by AI Atlas. All rights belong to the original author.