一个 Agent 需要知道目标、当前依据、行动边界、历史入口、未知信息的取得路径，以及离完成还差什么。但这些要求不会自己变成一份好的模型输入。
真正落地时，要回答更具体的问题：系统提示词里写什么？记忆在什么时候读？工具返回了一大张表，要不要每轮都带着？用户更新了条件，哪些材料继续有效？模型说“记住了”，我们如何知道真的保存了？
实现的主线，是在每次调用模型之前，从有效任务、获准材料和真实执行记录中构建当前工作视图；行动以后，再让新的观察和变更进入下一轮。 系统提示词提供稳定指导，记忆形成跨任务积累，工具结果与进度支撑当下推进。它们不是一个不断增长的字符串。
下面沿一个教学任务展开：用户比较基金 A、B 的费用与持仓重叠，要求结合上传的个人持仓生成带依据的报告，不执行调仓；研究中途上传新持仓，并明确要求以后基金研究报告不要默认压成短答。场景、输入别名和代码示例均用于教学，不是生产接口或一条实际运行记录。项目经验来自 ETF Agent 的已有实现与文章，不能将全文当作公开教学仓库的完整功能清单。
本篇不是一份必须逐项开启的机制清单。解释一个基金费用术语，直接回答可能就够；比较具体基金的费用与持仓，才按需读取条款、披露并计算；只有多轮修正、资料过长或中断恢复成为真实需求时，才考虑持久工作状态、压缩或子代理。先选择能完成任务的最小路径，再由可观察的失败决定是否增加机制。
一、先保留任务本身，再决定第一轮给什么
图 1：这是逻辑分区，不是一条 system 消息。发送前核对实际请求，不只看模板。 放大查看
任务刚开始，不要急着搜索，也不要先摘要用户要求。先让原始请求、明确限制、用户提交的材料和任务身份有可回到的记录。需要一个紧凑任务说明时，可以从这些原件构建，但不能让说明替代原件。
教学任务的有效要求可以表达为：
任务：比较基金 A、B 的费用与持仓重叠。
个人输入：用户本次上传的持仓快照 P1。
交付：一份可核对来源、口径和限制的报告。
限制：只研究，不执行调仓。
当前缺口：份额、适用费用、可比持仓披露尚待核对。
这不是预设工具执行计划。查到费用后可以先形成相应结论；没有完整持仓时可以扩大检索，也可以限定回答范围。用户原文中无法确定的部分，应保留疑问，而不是在生成任务说明时悄悄补成默认事实。
随后才选择第一轮输入。一个实用的划分是：
这里区分的是职责，不要求新增六张表或一个上下文总管理器。已有会话、工具、记忆、结果和运行记录可以继续各守自己的责任，在请求装配处形成一份一致的视图。
二、写系统提示词：把稳定指导与动态材料分开
一份有具体信号、又不硬编码研究路线的提示词
系统提示词要说明怎样工作，却不应替模型预写每一道题的完整步骤。下面是一份压缩过的教学模板，四个区块分别表达背景、行为原则、工具指引与输出要求：
<background_information>
你协助用户完成有依据的研究。当前目标、资料和能力由本轮输入提供。
你的任务是理解问题、选择研究路径、比较证据并组织结果。
</background_information>
<instructions>
围绕用户当前目标推进；区分目标、约束和可以修订的计划。
区分原始资料、计算结果、历史记忆、工作假设与先前回答。
资料不完整或存在冲突时，识别具体缺口及其影响，不补造事实。
用户修正条件后，重新判断受影响的工作，保留仍然适用的依据。
</instructions>
<tool_guidance>
根据实际提供的工具契约选择行动，不假设不存在的能力。
需要原先取得的材料时，使用受控回读；需要新事实时再获取或刷新。
缺的是用户选择时提出问题，不以无关检索代替。
保存、修改、遗忘等操作，以真实工具回执为准；结果未知就保留未知。
</tool_guidance>
<output_description>
按用户要求交付；关键判断带上适用条件和可核对依据。
区分已回答的部分、尚未解决的问题与下一步所需信息。
格式可以适配任务；本区不替代工具参数中的正式输出协议。
</output_description>
XML 在这里用于分区，不是权限机制，也不是必须使用的格式。使用 Markdown 标题也可以；应用仍须将可信指令、用户消息与工具资料放在各自正确的载体中，不能仅凭一层标签把外部文字升级为系统命令。
产品当前能力、用户姓名、持仓数据、今天的日期和旧记忆不应直接焊进这段稳定模板。它们来自不同来源，按任务装配。基金比较的份额与披露口径属于领域方法，可以放进相关 Skill；参数形状和工具用途应由工具契约清楚表达。
ETF Agent 将入口判断与持续研究的指导分开，研究方法也按既有入口或按需读取路径进入上下文。普通读取一份 Skill 正文，并不自动把正文重新注入为系统指令。这个选择的重点是来源与作用保持明确，而不是将每种文本都提升到最高优先级。[3]
用少量例子表达选择，不用边角清单替代判断
图 2：具体到能指导判断，不预写全部路线。用户、日期、持仓、记忆按轮装配。 放大查看
可以在方法说明中放一个典型示例：A 有完整持仓、B 只有前十大披露时，可以解释已知共同持仓，但不能称为完整重叠率；费用条款完整时，费用部分仍能独立回答。例子让“部分可回答”变得具体，而不要求模型匹配几十个固定分支。
例子里的基金和数值必须标成示范，不参与当前用户分析。几类有代表性的任务与反例，比把每次失败都追加成一条永久禁令更值得维护。Anthropic 的原文也建议从最小充分提示出发，依据实际失败补充清楚的指令和代表性示例，而不是堆积所有边角情况。[1]
修改时，先定位失败在哪里。如果内容根本没送到模型，修输入；如果工具职责不清，修工具说明；如果是本领域的方法问题，优先修相关 Skill；只有跨任务都成立的行为缺口，才值得进入稳定系统提示词。维护时可以对照修改前后的实际请求，检查指导是否重复、是否在正确阶段出现，再用同类任务比较效果。文本变短或结构检查通过，不能代替实际效果的比较。
三、准备首轮请求：先处理记忆准入，再装配实际可用能力
记忆先检查是否能读，再决定读哪些
假设用户此前保存过“回答默认简短”。本次要求深入比较时，不应不加区分地注入全部历史，更不能让这条旧偏好覆盖当前要求。
首轮可以先读取少量当前有效的偏好或回答约束，再按任务需要查找相关历史。临时模式则应在任何记忆库读取之前生效：既不召回，也不保存或自动学习；不能先读完，再让模型承诺不用。授权失败、读取不可用和库确实为空，也要保留不同结果。[4]
在候选选择上，先排除不属于当前范围、已失效或被遗忘的记录，再考虑与当前问题的相关性、适用时间和容量。关键词或向量检索只是候选发现方法，不是记录仍然有效的证明。少量明确偏好可以前置；长篇历史讨论更适合按需查找。
一个进入模型的记忆片段，至少要能说明它是什么背景。例如：
记忆背景：用户过去通常希望先给结论，再解释依据。
来源：用户明确提出的长期回答偏好。
用途：辅助组织表达；本次明确的深度要求优先适用。
本次未载入：与基金比较无关的其他历史讨论。
这是给模型的语义说明，不要求把所有存储元数据都塞进去。后台仍需要保留准确身份、修订、来源与有效状态，以便在内容改变或恢复时重新定位。
工具呈现既不能扩权，也不能隐藏仍然可走的路
图 3：临时模式在访问记忆库前生效。检索命中 ≠ 已发送 ≠ 正确采用。 放大查看
工具执行端验证当前调用是否获准，模型可见工具集则应反映本轮真实能力。两者需要一致，不能一边展示已不可用的工具，另一边让模型反复撞墙；也不能因为搜索失效，就移除读取已保存结果的合法入口。
保持工具名称、用途、参数和返回语义清楚，优先解决职责重叠，不必机械追求最少数量。能力较多时可以提供受控发现入口，发现后仍须正常准入。工具描述中的“可以查询”不是新的授权，目录中的文件名也不是已读证据。[3]
看一份首轮工作视图，而不是只看 Prompt 文件
对教学任务，首轮可见内容大致如下。它是逻辑分区，不是要求所有内容拼进一条 system 消息：
稳定指导：当前版本的系统提示词。
当前任务：比较 A、B；采用 P1；交付报告；不调仓。
有效记忆：先结论后依据；不覆盖本次详细分析要求。
方法入口：基金比较方法的简短说明，必要时读取正文。
资料入口：上传快照 P1、A/B 文件目录及其时点；目录不等于正文已读。
工具：本次获准且可用的读取、检索、计算、交付及必要记忆操作。
工作状态：费用与可比持仓问题尚待核实。
发送前应检查实际请求，而不只检查最初装配的对象。后续容量处理、扩展回调和模型适配，可能改变工具、材料和消息表示。关键问题是：原任务和修正还在不在，工具契约是否一致，新读取的内容有没有被变回一个空引用，记忆是否仍是获准使用的当前记录。
四、每次缺信息，先判断该回读过去，还是取得新观察
为资料保留两种入口：发现与深入读取
模型发现缺口后，需要能理解下一步去哪找。文档目录可以先返回产品、版本、披露时期、章节与页码；搜索结果提供候选来源；只有进一步读取才能获得所需正文。不要把标题命中写成“已核对报告全文”。
一个查费用的过程可以很短：确认份额，定位费用章节，读取适用条款，必要时查看脚注。持仓重叠则可能需要更多范围与结构化数据。工具应允许根据问题扩大读取，而不是固定“只读前三页”或每次整本返回。
这里也要分清问题类型。用户问“上次为什么那样判断”，先回到当时的报告、资料和计算；需要新一期披露，才去获取新版本；缺一个由已有数据可以算出的指标，则调用计算；缺用户选择，就提问。
把结果保存为可回读版本，避免一切都经过模型搬运
ETF Agent 的结果读取路径保留获准的结果版本，并通过 resultRef 与 read_tool_result 支持后续回读。下面使用教学别名解释它，而不是展示可直接调用的真实引用：
{
"resultAlias": "R_B_1",
"source": "基金 B 的某期持仓披露",
"dataAsOf": "2026-06-30",
"representation": "top-10-holdings",
"readStatus": "excerpt-read",
"selection": {
"fields": ["instrumentId", "weight"],
"offset": 0,
"limit": 10
},
"coverage": "仅覆盖来源披露的前十大持仓，不是完整基金持仓"
}
真实引用由宿主生成，并绑定原结果身份；回读要核对范围、版本和可读性。刷新应形成新版本，而不是静默改写 R_B_1 所指的材料。即使同名查询参数一样，也不能把新的外部结果当成原快照。[2]
大表格可以直接交给获准的计算工具。模型指定结果引用及选区，计算端读取原数据，返回结果、口径、输入身份和限制。模型需要核查异常记录时，再回读特定范围。减少的是无意义的复制，不是删除计算依据。
让覆盖信息一直跟到最终输入
图 4：resultRef 指向保存的版本，而非重查命令。回读 R1，不能偷偷换成 R2。 放大查看
曾有一次受控排查，50 条记录只剩 20 条进入实际模型请求。[2] 这说明“工具调用成功”不够：取回数量、过滤和分页、字节截断、目标记录身份，以及最终请求中的表示，都可能改变覆盖。
如果只拿到 B 的前十大，模型应知道当前能够计算的是已知共同持仓，不是完整重叠率。完整读取一份局部披露，仍不等于覆盖完整对象。对开放互联网搜索，更不能把返回若干条结果当成“相关资料已找全”。
当前刚读到的关键内容不应在装配时无声退出；较早结果可以缩成可精确回读的引用，但要保留选区、来源与覆盖。取得了、存好了、引用可用、正文送到了，是不同环节。 任何一个环节都不能替其他环节作证明。
五、新证据和用户修正到来后，更新的不只是聊天历史
假设研究已取得费用条款，并按持仓 P1 生成一个个人暴露计算。用户现在上传 P2，说“之前的持仓过时了，按这个算”。
把这句话追加到末尾，只完成了信息保存，还没有说明下一步该使用什么。一个更有用的更新是：保留用户修正与 P2 的身份，在当前视图中说明个人输入已变；原计算仍保留它采用 P1 的事实；把受影响的研究问题重新带到待处理状态。
代码可以记录确切输入版本与已知依赖，但哪些论证受影响，不一定能从一个通用规则机械判断。明确的领域计算可以由领域逻辑处理，更复杂的结论仍需要模型或审阅者重新检查；不能假定已经拥有一套自动识别全部语义影响的系统。
便签维护当前问题，回执维护执行事实
图 5：教学场景：只更新受影响的工作。保留旧结果身份，不把历史改成新事实。 放大查看
ETF Agent 用 researchNeeds 保存模型可修订的研究问题，并通过 set_research_needs 更新。下面仅示意它适合表达的内容：
费用：条款已取得，当前份额口径下可以作比较。
基金重叠：缺 B 的同期间完整披露，完整重叠率仍未确认。
个人暴露：用户改用 P2，原计算需要更新。
交付：先给有依据的费用结论，再整理最终报告。
模型标记一个问题已解决，是研究判断；工具返回可读结果、计算完成、报告保存，则各有真实记录。便签不能靠删除问题消除用户要求，也不能用“已保存”代替保存回执。[5]
更新便签也不必成为每轮强制动作。简单任务或状态未变化时，可以直接继续。重要的是在关键证据、冲突、用户修正或阶段变化后，后继请求能够获得有用进展，而不是一份已经过时的计划。
六、让记忆真正生效：写入、召回、纠正和遗忘要连到后续请求
记忆不只在任务结束时发生。显式“记住”“改正”“忘记”可能在对话任意时刻出现；可选的自动学习则可以沿合格回答的持久终态触发。两条路径不能混为“模型看过，就自动存了”。
显式保存：语义由模型理解，成功由持久操作回答
用户说“以后写基金研究报告不要默认压成短答，要讲清依据与口径”，表达了持续的回答偏好。用户只说“这次讲详细些”，则可以只影响当前回答。系统需要区分两者，不能因为一次任务中频繁出现某种写法就认定永久偏好。
ETF Agent 的显式路径由研究阶段在本账户范围内搜索已有记忆，再提议保存、修订或遗忘；宿主检查真实用户来源、目标版本和持久操作。其逻辑可以概括为：[4]
用户明确的持久请求
→ 搜索并定位已有记忆或确认可用的空结果
→ 提议新增、替代或遗忘
→ 检查范围、来源、目标修订及操作身份
→ 提交并取得回执
→ 根据回执报告结果
一个便于理解的存储示意如下。字段用于说明设计，不是项目原样接口：
{
"kind": "response_preference",
"scope": "本账户",
"family": "research-answer-depth",
"revision": 2,
"content": "基金研究报告应讲清依据与口径，不默认压成短答",
"origin": "explicit_user_request",
"sourceRef": "用户原消息 M2",
"status": "active",
"supersedes": { "revision": 1 }
}
记忆条目要能回答谁提出、适用在哪、当前是哪一版、为何仍有效。一次模型推断不能自行变成 user_confirmed；保存失败或结果未知时，也不能对用户说“以后都记住了”。本轮仍可按明确要求继续工作，但长期生效尚未确认。
自动学习：先是候选，不是悄悄改写系统提示词
图 6：保存、纠正、遗忘，都不是一句口头承诺。提交后回答失败，不回滚已提交记忆。 放大查看
普通研究中产生的经验，未必都是用户偏好。模型说“用户大概很保守”，可能只是自己的解释；一次成功任务采用了某个方法，也不等于以后每次都必须这样做。
可选自动学习应从获准的来源提出候选，保留语义判断与持久事实的区分。ETF Agent 的通用默认配置关闭自动语义学习，但保留显式记忆操作；观察模式可以保留候选而不自动生效，自动模式还依赖独立配置、来源与资源条件。[4]
工作笔记也不是自动学习的同义词。researchNeeds 负责本次研究的问题清单；跨独立任务的自主笔记可以记录项目探索与决策，但需要明确的保存和再读取路径，不能直接冒用用户偏好的身份。若要把经验沉淀为 Skill，应经过方法审阅与适用范围确认，而不是让模型自动追加全局系统指令。
召回：相关只是其中一道条件
下一次用户要求深入分析时，系统先验证当前记忆记录，再将相关修订放入实际请求。命中旧的 revision 1，并不意味着可以绕过 revision 2；相似度再高的被遗忘记录，也不能重新成为有效背景。
可以用一条实际消费链检查召回：合法候选 → 当前有效记录 → 本轮相关选择 → 容量装配 → 实际模型请求。每一步都可能减少内容，也可能失败。检索返回空、不获准访问、读取超时以及因容量未装入，不能统一伪装成“用户没有记忆”。
模型接收到了记忆，仍不代表它正确采用。可以另外检查回答是否遵循本次深度要求，但不要用“搜索命中次数”代替这项判断。ETF Agent 的使用记录按照实际派发统计，而不是把浏览、检索命中或历史重放都算成一次模型使用。[4]
纠正与遗忘：要阻止旧内容通过其他路径回来
纠正时保留修订关系，后续选择当前有效版本；遗忘时则使相应记忆失去未来使用资格。需要关注的不只有原记录：旧搜索结果、缓存、历史摘要、工作便签和晚到的学习任务，都可能携带它的派生内容。
在项目的生命周期设计里，记忆来源与依赖关系使后续投影能够检查这些派生材料，而不是仅从界面删掉一行。[4] 具体实现可能选择过滤、重新生成或暂不提供受影响材料；选择取决于系统能否可靠定位依赖，不能靠简单替换一个关键词保证遗忘完整。
同一研究的既有授权，也不意味着旧内容永远有效。运行授权与记忆的内容修订是不同事实：用户已获准使用本账户记忆，仍不应继续使用后来被纠正或遗忘的版本。新任务的权限则独立检查。
一个值得检查的场景是：遗忘操作已经提交，随后回答失败。下次进入时，应根据真实记忆状态继续排除该内容，不能因为聊天窗口没有收到“已忘记”就回滚操作。反过来，模型先说了“已忘记”，但写入未提交，也不能把那句话当成事实。
七、长任务减负：让旧内容退出，但保留继续工作的条件
图 7：教学场景：旧历史仍可存在。旧缓存、摘要和晚到任务不能使其复活。 放大查看
先区别存储、驻留和回读
系统能够保存很多数据，不代表模型应该始终携带它们。保存容量、自动放入请求的工作集容量、一次显式读取的容量，分别回答不同问题。不要为了减少自动注入，顺手把获准资料变成再也取不到的内容。
减负时可以先处理重复或较旧、且能够精确回读的工具结果。引用保留来源身份、版本、原选区与覆盖；本轮刚取得的关键材料、必要用户修正和正在使用的依据，需要在相应决策中可见。
在一轮真实研究观察里，Main 的 37 次原生请求中有 20 次显式读取，相同读取结果在后继消息中又累计出现 279 次。它们是重复携带，不是额外执行 279 次外部查询。后来对较旧的可回读结果采用更小的工作集，在受控请求对照中减少了超过 30,000 字节；这只能证明装配变化，真实费用还要考虑缓存、回读及最终任务效果。[2]
因此，不能只问单次请求缩短多少。过早移走有用内容，可能导致反复回读；保留全部内容又可能让历史噪声持续累积。应对照同一任务的整体耗时、质量与调用开销选择策略。
确实需要摘要时，把它当作有损工作表示
图 8：先保留回到原件的路径。摘要描述工作，不改写证据与效果。 放大查看
语义压缩与原文回读不同。摘要会选择和改写信息，所以应限定它能够处理什么、必须保留什么，以及失败时如何回到原表示。
ETF Agent 的同 Run Main 压缩选择有原始依据的已结算旧工作组，保留近期轮次与用户原文，结果引用仍来自真实回执。这里的摘要输入主要是已经封存的模型工具提议，而不是宣称能够概括全部原文与隐含思考。摘要描述旧工作线索，不能凭一句“已经查完”证明执行成功。[6]
采用前还要把摘要和引用放回完整请求，检查实际净收益与输入身份；不够省、放不下或来源不匹配，就保留原表示或按已有容量路径处理。生成摘要本身消耗资源，不能在没有收益时无限重试。
一旦采用，记录它替代的区间、来源、版本与请求身份，让后续重建能对回原件。同一次研究的压缩与跨研究会话历史整理也要分开：前者服务当前执行，后者提供下一次任务的历史背景，不重新发起旧任务，也不自动把旧便签变成本轮进度。
子代理和恢复，都是重新构建合适的局部视图
委派费用核对时，提供基金身份、份额、局部目标、必要限制和获准资料即可，不必复制全部私人持仓、长期记忆和其他探索历史。子代理带回判断、适用条件、依据与未决项，主代理再把它接回整体任务；一句“费用更低”不足以完成交接。[7]
恢复时也不是只发送一段摘要加“继续”。应读取原任务及修正、有效输入版本、工作状态、已提交结果和资源事实，并恢复系统支持的接续点。未知效果不能自动重做，已提交的记忆操作和报告也不能凭本地缺少成功提示就被撤回。
这些方式改变的是模型当下携带的材料。它们都不应该重新发明用户目标、原始证据或已经发生的操作。
子代理、压缩和回读：从可选方案开始
对于两只基金的费率和持仓重叠比较，默认方案是单 Main 加必要工具。若短的原始材料已经足以回答问题，就不必为了「治理」将它外置后再回读。只有独立资料确实造成上下文干扰，或并行探索产生可比较的收益，才考虑子代理：它会新增模型调用、任务交接、结果身份和失败路径。压缩也不能只看一轮 Token 减少，还要算摘要调用、后继回读和错误恢复的成本。
八、沿着一项任务检查，而不是给每个模块单独打勾
图 9：局部视图更小，关键联系不能消失。核对实际输入、真实效果与最终交付。 放大查看
一套机制都存在，仍然可能在交接处失效。可以用下面几类输入，在实际请求、工具回执和最后报告之间做对照：
这些检查有不同性质。输入身份、范围、版本与操作回执可以进行确定性核对；有没有理解费用例外、是否真正回答用户问题，则需要模型或人工的独立评审。不要用前一种检查通过替代后一种。
观测也不必复制所有私人原文。普通记录可以保留请求身份、组成、版本、容量、表示类型和省略原因；需要查看具体内容时走受控读取。对于无法取得的观测，保留缺口，不能从“没有日志”推断“没有发生”。
最终交付应回答开头那项任务：采用了哪份个人输入，费用按什么口径比较，持仓重叠覆盖到哪里，哪些问题没有解决，以及报告是否真的可取得。记忆帮助表达和接续，系统提示词帮助稳定判断，二者都不能代替这份成果本身。
把任务质量与机制成本放在一起比较
可以保持用户问题、基金资料版本、模型配置和评价标准一致，比较基础 Main 与按需启用回读、压缩、子代理的候选方案。既检查报告是否回答原题、数字及来源是否正确、披露范围是否诚实、报告是否交付，也记录调用次数、Token、延迟和失败类型。没有足够的真实配对，就只能提出设计假设，不能声称多专家更好或成本已经显著下降。
还要观察正常路径税：高级能力没有触发时，普通请求是否仍携带额外的工具说明、状态与协调逻辑？如果长期增加负担，却没有可观察的任务收益，就应考虑合并、按需激活或退出。授权、来源与防止副作用重复等硬约束仍须保留，不能被简单的成本模型抵消。
结语：把做法放回同一条任务链
接到任务，保存原要求，构建稳定指导与当前目标；在允许范围内选择有效记忆，呈现可用工具与资料入口；围绕缺口发现、读取和计算；让新证据和用户修正改变后继输入与进度；需要长期积累时完成真实记忆操作；内容增多后有选择地减负，分工和恢复都继续沿原事实接续。
这就是系统提示词、记忆、检索、结果引用、便签、压缩和子代理共同工作的方式。落地的重点不是把它们逐项装齐，而是检查每一次信息交接是否仍然支持同一个有效目标。
从下一次实际模型请求开始检查：它知道该做什么吗，拿到了所需依据吗，能够找回过去和探索未知吗，记忆与进展还是当前有效的吗？ 能把这些问题对回真实输入和真实结果，上下文治理才从一组理念变成可维护的工程。
参考与延伸
[1] Anthropic（2025-09-29），Effective context engineering for AI agents。系统提示词、代表性示例及上下文工程的参考；本文中的模板是教学示例，不是原文摘录。
[2] 工具查到的数据，为什么不一定要给大模型看？。版本化结果、覆盖、回读及本文引用的两类输入观察。
[3] Agent 有这个工具，模型为什么还是看不到？与Model、Runtime、MCP、Skill 的责任边界。方法、工具与最终请求的职责关系。
[4] Agent Memory 的完整生命周期：从记住偏好到安全遗忘。显式操作、可选学习、当前修订、派发消费与遗忘边界。
[5] 为什么 AI Agent 也需要《思考，快与慢》？。可修订的研究便签与任务路由。
[6] 同 Run 上下文压缩与跨 Run 会话历史压缩。两种历史处理的不同输入和职责。
[7] Subagent：上下文隔离与专业化。局部视图与研究交接。
开源工程实践见 Open ETF Agent。若这些文章对你有帮助，欢迎 Star 关注；公开实现的实际范围以仓库说明为准。
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。
An Agent needs to know the target, the current basis, the operational boundary, the historical entrance, the path to the unknown information, and what is missing. But these requirements do not themselves become a good model input.
When you actually land, answer the more specific question: What is written in the system prompt? When do you read the memories? The tool returned a big watch. Do you want to carry it with you? Users have updated conditions and what materials remain valid? The model says, "Remember." How do we know it's actually saved?
The main line to be achieved is to construct the current working view from valid missions, approved materials and authentic execution records before each call to the model; after the operation, new observations and changes will enter the next round. The system prompt provides stable guidance, memory builds across missions, tool results and progress underpins current progress. They're not a growing string.
The following is a teaching mission: the cost of the User Comparison Fund A and B overlaps with the holding stock, requiring that the feed be combined with the basis of the report on the production of the individual hold hold; and that the new hold be uploaded midway to the study, with a clear request that future Fund studies not be tacitly reduced to a short response. The scene, the entry of aliases and examples of codes are used for teaching, not for production interfaces or a physical running record. Project experience from ETF Agent has already been realized and articles, and the full text cannot be used as a complete list of functions for an open teaching warehouse.
This section is not a list of mechanisms that must be opened on a case-by-case basis. An explanation of a fund ' s cost terminology and a direct response may be sufficient; a comparison of the costs and holdings of a specific fund would be required to read, disclose and calculate the terms; and a permanent working situation, compression or sub-agent would be considered only if multiple rounds of amendment, lengthy information or interrupted recovery became a real need. First, the smallest path through which the mission can be accomplished is chosen, and then the observable failure decides whether or not to add mechanisms.
I. Retention of the mandate itself before deciding what to give in the first round
Figure 1: This is a logical partition, not a system message. Checks actual requests before sending, and not just looking at templates. Zoom in
The mission has just begun, so don't rush to search or start with summary user requests. The original request, the clear limits, the user ' s submissions and the identity of the mission are first recorded. When a compact mission statement is required, it can be constructed from those originals, but it cannot be replaced.
Effective requirements for teaching tasks can be expressed as:
• Mandate: ICF A and B cost overlaps with hold.
Personal input: User's current upload of warehouseshot P1.
Delivery: a report that can verify sources, calibres and limitations.
Limitations: research only, no reconciliations implemented.
Current shortfall: share, applicable cost, comparable warehouse disclosure pending reconciliation.
This is not a pre-set tool implementation plan. Costs can be identified before conclusions can be drawn; access can be expanded without complete hold, or the scope of answers limited. The undetermined part of the user's original text should be retained in doubt, rather than quietly added to the default fact when the mission statement was generated.
The first round of input is selected later. A useful distinction is:
This distinction is made between duties and does not require the addition of six tables or a context master manager. Meetings, tools, memories, results and operational records are available to maintain their respective responsibilities and to create a consistent view at the request assembly.
II. Writing system prompt: Separation of stabilization guidance from dynamic material
A prompt with specific signals and uncodified lines.
The system prompt is intended to explain how it works, but should not prewrite the complete steps of each issue for the model. The following is a compressed teaching template, with four blocks expressing background, principles of behaviour, tools guidance and output requirements:
<background_information>
You're helping the users to complete the evidence. Current objectives, information and capabilities are provided through the current round of inputs.
Your task is to understand the problem, to choose the path, to compare the evidence and to organize the results.
</background_information>
<instructions>
Drive around current user objectives; distinguish between objectives, constraints and plans that can be revised.
Distinguishing between source data, calculations, historical memory, working assumptions and previous responses.
When information is incomplete or there is conflict, specific gaps and their impact are identified and do not make up for the facts.
After the user has amended the conditions, the affected work is rejudged and the basis on which it remains applicable is retained.
</instructions>
<tool_guidance>
The selection of actions based on the actual availability of the tool compact does not assume non-existent capabilities.
Use controlled rereading when previously acquired material is required; re-acquisition or refreshing is required when new facts are required.
What is missing is that questions are raised when the user is selected and not replaced by unrelated searches.
Save, modify, forget, etc., depending on the true instrument; the result is unknown.
</tool_guidance>
<output_description>
Delivery according to user requirements; key judgement with applicable conditions and verifiable basis.
Distinguishing between the part that has been answered, the outstanding issues and the information required by next step.
Format to fit tasks; this sector does not replace formal output protocols in tool parameters.
</output_description>
XML is used here for partitions, not the permission mechanism, or the format that must be used. The use of the Markdown title is also possible; the application must still place credible commands, user messages and tool materials in the correct vehicle, and cannot upgrade external text to a system command by a single label.
The current capacity of the product, the name of the user, the holding data, today ' s date and the old memory should not be welded directly into this stable template. They come from different sources and are organized according to mission. The Fund ' s comparative share and the disclosure of calibres as an area approach could be placed in the relevant Skill; the shape of the parameters and the utility of the tool should be clearly expressed in the tool compact.
ETF Agent separates the entry judgement from the guidance of the ongoing study, which also enters context by the established entry or on-demand reading path. Reads a regular Skill text and does not automatically re-injecte it into a system command. This option focuses on clarity of origin and role, rather than raising each text to its highest priority. [3]
The choice is expressed in a small number of examples and the margin list is not used to substitute judgement
Figure 2: Specific enough to guide judgement and not predetermine all routes. Users, dates, hold-ups, memories are assembled by wheel. Zoom in
A typical example could be given in the description of the methodology: A with a full hold and B with only the top 10 disclosures could explain the known common hold, but not the full overlap rate; when the cost clause was complete, the cost portion could still be answered independently. Examples make “partly answerable” specific without requiring models to match dozens of fixed branches.
The funds and values in the examples must be labelled as demonstration and not involved in current user analysis. Several categories of representative mandates and counter-examples are more worth preserving than adding a permanent ban to each failure. The original text of Anthropic also suggests adding clear examples of directives and representation based on actual failures, rather than stacking all the corners. [1]
When changing, locate where the failure is first. If the content is not delivered to the model, the input is repaired; if the duties of the tool are unclear, the tool is described; if it is a methodological issue in this field, priority is given to repairing Skill; only if a behavioural gap has been created across the mission is worth entering the stabilization system prompt. Maintenance can be performed by checking whether the guidance is duplicated or not at the right stage, against actual requests before and after the modification, and then comparing the effects with the same tasks. Shorter texts or structural checks are passed and cannot substitute for comparison of actual effects.
III. Preparation of the first round of requests: processing of memory access and assembly of actual availability
Let's see if we can read it and decide what to read.
Assuming that the user has previously saved the "Default Shortness of Answers". This request for in-depth comparisons should not be indiscriminately injected into history, let alone allow this old preference to cover current demands.
The first round allows for a small number of current valid preferences or answer constraints, followed by a search for the relevant history as required by the mission. The ad hoc model should take effect before any memory library is read: neither recall nor retention or automatic learning; it cannot be read before the model is committed to non-use. The failure of the authorization, the lack of access and the fact that the library is empty also require the retention of different results. [4]
In the selection of candidates, records that are not in the current context, have lapsed or forgotten are excluded before the relevance, timing and volume of application to the current issue is considered. Keywords or vector searches are only candidate methods and are not proof that the record is still valid. A small number of explicit preferences can be pre-empted; lengthy historical discussions are more appropriate for searching on demand.
A memory clip into the model, at least to show what it is. For example:
Memory background: Users have in the past often wished to draw conclusions before explaining the basis.
Source: Long-term preferences clearly expressed by users.
Purpose: Expressions by supporting organizations; this clear-cut depth requires priority application.
Not included: Other historical discussions not relevant to the comparison of the Fund.
This is a synonym for the model and does not require all stored metadata to be inserted. There is still a need to maintain accurate identification, revision, origin and validity in order to reposition when content changes or restores.
Tools can neither be expanded nor hidden from the path that remains.
Figure 3: The provisional model takes effect before accessing the memory library. Retrieval hits the gills have sent the gills for correct use. Zoom in
The tool implementation end validates whether the current call is approved and the model visible tool set should reflect the real capacity of the current round. The two need to be consistent, not to display tools that are no longer available on the one hand, to allow models to hit walls repeatedly on the other; nor to remove the legitimate entry point for reading the saved results because the search has failed.
Maintain clarity about the name of the tool, its use, parameters and the semantics of return, giving priority to addressing the overlapping of responsibilities without having to pursue the minimum number mechanically. Controlled detection portals can be provided when greater capacity is available, and regular access is required after detection. The " searchable " in the tool description is not a new mandate, nor is the filename in the directory readable evidence. [3]
See a first round of work views instead of just Prompt files
For teaching tasks, the first round can be summarized as follows. It's a logical partition, not a system message:
Steady Guidance: Current version of system prompt.
Current tasks: comparison A, B; introduction of P1; delivery report; no encumbrance.
Effective memory: based on conclusions; does not cover this detailed analytical requirement.
Methodology portal: A brief description of the Fund ' s comparison methodology, reading the main text as necessary.
Information portal: Upload snapshots P1, A/B file directories and their time points; directories are not the same as read text.
Tools: This allowed and available reading, retrieval, computing, delivery and necessary memory operations.
Status of operations: Costs and comparable holding are yet to be verified.
The actual request should be checked before dispatch, and not only the initial assembly. Follow-up capacity processing, extended back-to-back and model adaptations may change tools, materials and messages. The key questions are whether the original mandate and amendments are still in place, whether the instrument contract is consistent, whether the newly read content has been turned back into an empty reference and whether the memory remains the current record of authorized use.
IV. Whether to read the past or get new observations each time information is missing
Keep two portals for information: discovery and in-depth reading
When the model finds a gap, it needs to understand where next step should be found. The document directory can first return the product, version, disclosure period, chapter and page numbers; the search results provide a candidate source; only further reading can obtain the required text. Do not write the title to read “the full verified report”.
A process of checking costs could be short: recognition of shares, section on locator costs, reading of applicable provisions and looking at footnotes if necessary. Overlaying may require additional scope and structured data. The tool should allow for expanded reading based on the problem rather than a fixed " first three pages only " or a full return.
Here too, the type of problem should be distinguished. Users asked “why was that last judgement” to go back to the report, information and calculations at the time; a new issue of disclosure was required before a new version could be obtained; an indicator from which data were available was missing, then the calculation was called; and a user selection was missing and questions were asked.
Save the result as a rereadable version and avoid all being moved through the model.
The result of ETF Agent retains the approved version of the result and supports subsequent re-reading with read tool result. The teaching aliases are used to explain it, rather than to demonstrate the true references that can be called directly:
{
"resultAlias": "R_B_1",
"source": "A certain period of Fund B disclosure,"
"dataAsOf": "2026-06-30",
"representation": "top-10-holdings",
"readStatus": "excerpt-read",
"selection": {
"fields": ["instrumentId", "weight"],
"offset": 0,
"limit": 10
},
"Coverage": "Only the top 10 warehouses, not the whole fund. "
}
True references are generated by the host and the original result is identified; reread to check scope, version and readability. The update should be a new version, not a silent rewriting of the material referred to in R B 1. Even with the same search parameters, new external results cannot be considered as original snapshots. [2]
Large tables can be handed over directly to approved computing tools. Model results specify the reference to the constituency, calculate the end to read the original data, return the results, calibre, input identity and limits. The model needs to reread a specific range when verifying an abnormal record. The reduction is meaningless reproduction, not the deletion of the basis of calculation.
Let the overwhelm information follow through to final input
Figure 4: Result Ref points to a saved version, not to a recheck command. Read R1, not sneaking into R2. Zoom in
There was one controlled check, and there were only 20 requests from 50 records to enter the actual model. [2] This means that “tool call has succeeded” is not sufficient: retrieval of quantity, filtering and page break, byte cut-off, target-recording identity, and representation in the final request may change coverage.
If only the top 10 of B is available, the model should know that the current calculation is known to be a shared hold, not a full overlap rate. Full reading of a partial disclosure is still not equivalent to covering the full object. For an open Internet search, it is even less possible to consider the return of a number of results as “relevant information has been found”.
Key elements currently being read should not be withdrawn silently at assembly; earlier results could be reduced to precise rereadable references, but the constituency, source and coverage should be preserved. There are different links between the acquisition, storage, use of references and delivery of texts. None of these links can prove otherwise.
V. New evidence and user corrections update more than just chat history
Assuming that a cost clause has been obtained by the study, the calculation is based on holding P1 generating a personal exposure. The user is now uploading P2 and says, "The previous holdout is out of date, at this rate".
Add this sentence to the end and only information storage has been completed, with no indication of what next step should be used. A more useful update would be to retain the identity of the user-modified and P2 and to state in the current view that the individual input has changed; to retain the fact that it uses P1 in the original calculation; and to bring the issue of the affected research back to the status of the pending.
The code can record the exact input version and the known reliance, but what arguments are affected may not necessarily be judged by a generic rule machine. Clear field calculations can be dealt with by the logic of the field, and more complex conclusions still need to be rechecked by models or reviewers; it cannot be assumed that a system that automatically recognizes the full semantic impact is in place.
Sign to maintain the current problem and back to enforce the facts
Figure 5: Teaching scenes: only the jobs affected are updated. Retain the old outcome and not change history to a new reality. Zoom in
ETF Agent saves model-modifiable research issues with searchNeeds and updates them by set research needs. The following is just an indication of what it is appropriate to express:
Cost: Provisions have been made and can be compared under current shares.
Overlap of funds: Full disclosure of gap B over the same period and full overlap rate not confirmed.
Personal exposure: Users switch to P2 and original calculations need to be updated.
Delivery: a valid cost conclusion followed by final report.
Model tags are a problem that is solved by research judgement; the tool returns the readable results, the calculation is completed, the report is kept, and there are real records. Signing cannot eliminate user requirements by deleting questions, nor can it be replaced by “saved” to save a reply. [5]
The updates do not necessarily become mandatory per round. A simple task or state may continue directly if it remains unchanged. It is important that, after key evidence, conflict, user correction or stage change, subsequent requests achieve useful progress rather than an outdated plan.
VI. Making memories real: writing, recalling, correcting and forgetting are linked to follow-up requests
Memory does not occur only at the end of the mission. Visible “remember” “correct” “forget” may appear at any point in the dialogue; optional automatic learning can be triggered by a lasting finality along the lines of a qualified answer. The two paths cannot be mixed into "models are seen and automatically stored."
Visible saving: semantics understood by models, successfully answered by persistent operations
Users expressed a continuing preference for a “subsequent Fund study without defaulting on short answers, with clear justification and calibre”. Users say only “a little more detail this time” and can influence only the current answer. The system needed to distinguish between the two, and a permanent preference could not be established because of the frequent occurrence of a writing in a mission.
ETF Agent's visible path is searched for existing memory within this account at the research stage and proposes to be preserved, revised or forgotten; the host checks the real user source, target version and ongoing operation. The logic can be summarized as follows:[4]
Clear and sustained user request
Search and locate existing memories or confirm available empty results
• Proposed new, substitute or forgotten
• Check scope, source, target revision and operational identity
• Submit and obtain a reply
According to the report,
An understandable storage indication is as follows. field is used to describe the design, not the project interface:
{
"kind": "response_preference",
"scope": "this account,"
"family": "research-answer-depth",
"revision": 2,
"Content": "The Fund's research should be based and calibrated and not default to short answers."
"origin": "explicit_user_request",
"sourceref": "User message M2"
"status": "active",
"supersedes": { "revision": 1 }
}
A memory entry should answer who proposed it, where it was applied, which version it was currently, and why it was still valid. A model infers that it cannot become itself a user confird; nor can it say to the user, “Be remembered later”, when the preservation fails or the results are unknown. The current round may continue as expressly required, but its validity has not yet been confirmed.
Automatic learning: First candidate, not quietly rewrite system prompt
Figure 6: Preservation, correction, oblivion is not an oral commitment. The response after submission failed and no roll-back was submitted to memory. Zoom in
Experiences arising from general studies are not necessarily user preferences. The model says that “users are probably conservative” and may be a self-depreciation; a successful mission uses a method and is not the same as having to do so every time.
Optional automatic learning should be presented from an approved source, leaving a distinction between semantic judgement and enduring facts. The EFAgent 's Universal Default Configuration closes automatic semantic learning but retains a visible memory operation; the observation mode retains a candidate without automatic validity and relies on independent configuration, source and resource conditions. [4]
Work notes are not synonymous with automatic learning. ResearchNeeds is responsible for the list of issues for this study; autonomous notes on cross-independent missions can record project exploration and decision-making, but require clear preservation and re-reading paths that do not directly use user preferences. If experience is to sink to Skill, it should be subject to methodological review and application confirmation, rather than automatically adding global system instructions to the model.
RETURN: It's only one condition.
The next time the user requests an in-depth analysis, the system validates the current memory records before putting the relevant revisions into the actual request. The old vision 1 does not mean that it can be bypassed; the forgotten record, which is more similar, cannot be put back into a valid context.
A physical consumption chain can be used to check for recall: a legal candidate → the current valid record → the relevant selection for this round → capacity assembly → the actual model request. Each step may reduce content or fail. Retrieval is empty, access is denied, timeout is read and, because capacity is not loaded, it is not uniformly disguised as “user without memory”.
The fact that the model received memories still does not mean that it was correctly adopted. A separate check can be made of whether this Depth requirement is followed, but not of the “survey hits” as a substitute for this determination. The records of the use of ETF Agent are based on actual distribution, not on browsing, retrieving or replaying as a model. [4]
Correcting and Forgetting: To prevent old content from coming back through other paths
The correction retains the modified relationship and the subsequent selection of the current valid version; the forgotten memory loses its eligibility for future use. Not only the old search results, cache, historical summary, job signings and late-time study assignments may be of interest.
In the life-cycle design of the project, the relationship between memory sources and dependency allows subsequent projection to examine these derivatives rather than simply deleting a line from the interface. [4] Concrete realization of the possible choice of filtering, regeneration or withholding of affected material; selection depends on the reliability of the system ' s positioning, and a simple replacement of a keyword cannot guarantee the completeness of the forgotten.
The mandate for the same study does not mean that the old content will remain valid forever. The revision of operating authorizations and memories is a different fact: users have been granted access to the account ' s memory and should not continue to use later corrected or forgotten versions. The new mission permission is independently checked.
One scene worth checking is that the forgotten operation was submitted and the answer failed. The content should continue to be excluded from the next entry according to the state of real memory and should not be rolled back because the chat window does not receive "forget". The model, in turn, begins by saying “forget it”, but does not include the sentence, nor can it be taken as fact.
VII. Longer assignment minus negative: withdrawal of old content but retention of conditions for continuation of work
Figure 7: Teaching landscape: Old history still exists. Old cache, summary and late missions cannot revive them. Zoom in
Distinguishing storage, presence and reading back first
The system can save many data, which does not mean that the model should always carry them. Saves the capacity, automatically inserts the work volume of the request, and the capacity of a single visible reading, and answers different questions. In order to reduce automatic injection, the approved information should not be turned into something no longer available.
The reduction can first address duplicate or older tools that can be accurately read back. Quoting retention of source identity, version, original constituency and coverage; key material newly acquired, necessary user amendments and the basis for use in this round need to be found in the corresponding decision-making.
In a series of genuine research observations, Main's 37 original requests were read 20 times, and the same reading cumulatively appeared 279 times in subsequent messages. They are duplicates, not extra 279 external queries. Later, a smaller set of jobs was introduced for older reversible results, which reduced the comparison of controlled requests by more than 30,000 byte; this only proved a change in assembly, with real costs taking into account the effects of cache, rereading and the final mission. [2]
Therefore, it is not possible to ask only how much shorter a single request is. The premature removal of useful content may lead to re-reading; retaining all content may in turn allow historical noise to accumulate. Selective strategies should be applied against the overall time-consuming, quality and mobilization costs of the same mission.
When summary is really needed, it's a sign that it's bad work.
Figure 8: Retain the path to the original. summary describes the work and does not rewrite the evidence and effects. Zoom in
Semantic compression differs from the original reading. summary will select and rewrite information, so it should be limited to what it can handle, what it has to keep, and how it returns to its original expression when it fails.
The ETF Agent's same Run Main compression selection was originally based on a closed old working group that retained the recent rotation with the user's original text, and the result was still a true reply. The summary input here is mainly a proposal for a stored model tool, rather than a statement that it can summarize all the original language and implicit thinking. summary describes the old lines of work and cannot prove the success of the implementation with a “check-out”. [6]
summary and the reference are also to be returned to the full request, to check the actual net proceeds and the input identity; to be saved, left behind or not matched by the source, to be retained in the original expression or to be processed by the available capacity path. The generation of summary itself consumes resources and cannot be unlimited without the proceeds of retry.
Once adopted, it will record its substitute compartments, sources, versions and requests, so that subsequent reconstruction can be returned to the original. The compression of the same study is also separated from the cross-research session history: the former serves the current implementation, while the latter provides the historical background for the next mission, does not re-launch the old mission and does not automatically turn the old signer into the current round of progress.
Sub-agents and restorations are reconstructing suitable local views
The verification of the assignment costs is sufficient to provide the identity of the fund, its share, its local objectives, the necessary restrictions and the information to be approved, without the need to copy all private holdings, long-term memory and other exploratory history. The sub-agent takes the judgement, applies the conditions, relies on the outstanding items, and the main agent takes it back to the overall task; the phrase “lower costs” is not sufficient to complete the handover. [7]
It was not only a summary plus “continue” sent when it was restored. The original mandate should be read and amended, the valid input version, the working state, the results submitted and the fact of resources, and the system support should be restored. Unknown effects cannot be automatically reworked, nor can submitted memory operations and reports be withdrawn in the absence of local success indicators.
These changes are the material that the model is carrying right now. Neither should they reinvent user targets, original evidence or operations that have taken place.
Sub-agents, compression and rereading: Start with options
For comparison of rates and holdovers between the two funds, the default scheme is a single Main plus necessary tool. If the short source material is sufficient to answer the question, there is no need to re-read it for "governance". Sub-agents are considered only when independent information does cause context interference, or when parallel exploration produces comparable benefits: it will add models to call, task transfer, outcome identity and failure path. Compression cannot be reduced by just one round of Token, but also costs summary call, follow-up reading and recovery of errors.
viii. Check along a mission rather than ticking each module separately
Figure 9: The local view is smaller and the key link cannot be lost. Check the actual input, the true effects and the final delivery. Zoom in
A set of mechanisms exists and may still fail at the junction. The following types of input can be used to compare actual requests, tool responses and final reports:
These checks are of a different nature. The entry of identity, scope, version and operational response can be checked for certainty; whether the cost exception is understood or whether the user question is actually answered requires a model or manual independent review. Do not replace the latter with the previous inspection.
It is also not necessary to copy all private originals. The ordinary record retains the requested identity, composition, version, capacity, expression type and omission; it is subject to controlled reading when specific content needs to be viewed. For observations that cannot be obtained, the gap cannot be inferred from the “no log” that “does not occur”.
The final delivery should answer the initial task: which individual input was used, what caliber was the cost compared, where the silo overlap covered, which issues remained unresolved and whether the report was really available. Memory helps to express and continue, and the system prompt helps stabilize judgment, neither of which can replace the outcome itself.
Compare task quality with mechanism cost
User issues, Fund information versions, model configuration and evaluation criteria can be kept in line, comparing the basis of Main with the needs-based rereading, compression, sub-agent candidate. The report is checked for answers to the original questions, the correct numbers and sources, the integrity of the scope of disclosure, the delivery of the report, the number of calls, Token, the type of delay and failure. In the absence of a true match, only design assumptions can be made, and it cannot be claimed that multiple experts are better or that costs have decreased significantly.
It is also important to observe the normal path tax: do ordinary requests still carry additional tool descriptions, status and coordination logic when the advanced capability is not triggered? Integration, activation as required or withdrawal should be considered if the burden increases over a long period of time and there are no observable mission gains. Hard constraints such as authorization, source and prevention of duplication of side effects must remain and cannot be offset by simple cost models.
Conclusion: Returning the practice to the same task chain
Mandates to preserve original requirements and build stable guidance and current targets; to select effective memory to the extent permitted, presenting available tools and access to information; to detect, read and calculate around gaps; to allow new evidence and users to modify subsequent input and progress; to complete real memory operations as long-term accumulation is required; to selectively reduce content and to maintain division of labour and recovery as it continues.
This is how the system prompt, memory, retrieval, results reference, signage, compression and sub-agents work together. The focus of landings is not to assemble them item by item, but to check whether each transfer of information still supports the same effective goal.
Starting with the next physical model request: Does it know what to do, has it got the required basis, can it recover the past and explore the unknown, and is memory and progress still valid? The governance of context has evolved from a set of concepts to a work that can be maintained if these issues are translated back into reality and true results.
References and extensions
[1] Anthropic (2025-09-29), Effective contact engineering for AI treaties. The system prompt, the representative example and the reference for the context project; the template in this paper is a teaching example and not an extract from the original text.
[2] Data found in the tool, why not show LLM? Version results, cover, read back and two types of input observations cited here.
[3] Agent has this tool. Why is the model still invisible? The border of responsibility with Model, Runtme, MCP, Skill. The relationship between methodology, tools and responsibilities for final requests.
[4] Agent Memoory's complete life cycle: from remembering preferences to safe forgetting. Visible operations, optional learning, current revision, distribution consumption and forgotten borders.
Why does AI Agent need to think, fast and slow? The reversible studies sign off with the routing of tasks.
[6] Compression with Run context and cross Run session history compression. Different inputs and responsibilities from two historical processes.
[7] Subagent: context isolation and specialization. Local views intersect with research.
See Open ETF Engineering Practice. If these articles are helpful to you, Star ' s attention is welcome; the actual scope of public realization is based on the description of the warehouse.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.