六层的价值，不是把模型套进一条固定流水线，而是分清每次交接：谁提出工作、谁判断下一步、谁真正执行，以及哪份结果已经交到用户手中。 本文沿一次请求解释当前实现。六层是阅读地图，不是六个进程；跨层参与也不会改变各模块对权限、事实和效果的责任。
一、一次用户提问，为什么需要六个环节
下面用一个教学场景贯穿主链：用户在聊天框输入“分析我的基金组合，看看行业集中度和汇率风险”。它用于解释调用关系，不是一份投资结论或真实运行成绩。
从界面上看，只发生了一件事：发送消息，然后收到回答。从系统内部看，至少要解决六个不同问题：工作由什么输入或事件触发；该触发归属哪个用户与 Run、允许使用什么上下文；模型准备如何处理；是否需要调用工具；哪些内容真正获准发布；用户界面如何显示或恢复它。
读这张图，可以像追踪“输入 URL 到页面渲染”一样：先沿请求前进，再放大有分支和回环的地方。 普通概念问题可以直接回答，资料不足时进入研究；后台研究还在运行时，新消息又可能只是补充条件。这些情况共用底层能力，却不是同一条固定路径。
先不看 47 个模块：60 秒读懂六层
想象同事发来一句话：“帮我比较两份基金报告，告诉我差别，并保存下来。”从他点击发送到收到报告，可以先只记六件事：
如果第二份报告没读到，助手可以交代已经确认的部分，但不能把缺失的比较说成已完成。反过来，保存了报告也不意味着页面一定已经展示它。
下面的大图是技术索引，不是阅读前置考试。 第一次阅读可以先跳过 47 个模块名，只顺着这六个问题往下看；需要排查某一步时再点开对应模块。六层也不是每次请求都必须走完的固定审批流程。
交互导航： 技术站原文中的全景图支持点击 Owner 标签、全屏和缩放；也可通过 Owner 详情 查阅对应模块。
打开可放大的 SVG 全景图：9 个能力分组、47 个独立模块标签；研究工具与金融工具、身份与授权分别合列。
图中六条横带对应六个请求处理环节；每条彩色竖柱是为了阅读而聚合的一组能力。每个实线标签代表独立 Owner 模块；浅色虚线框代表处理动作或使用说明，并在框内标明“来源：对应 Owner”。顶部的虚线列标题只是阅读分组，并非单一模块。 纵向分组可以跨层、也可以断开，但不会因此增加授权或控制权。横向分层与纵向能力并不是彼此排斥的分类：一个 Owner 可以参与多层，但其事实归属不会因此复制。
这里的 L6 不是说界面只在最后出现。发送按钮和模型选择也发生在 L6，但 L1 不只接收聊天消息：已确认的定时 Task 也可触发 Research。随后 L2 对触发关联的身份、Run 和权限进行确定性核验。六层描述主链，不意味着每次都重新创建 Run。
二、L1 任务输入与触发：Agent 不一定从聊天框开始
把 L1 理解为任务起点，而不是 HTTP 认证层。用户点击“发送”是最常见的入口；用户主动要求生成周报、点击“今日研究”，或者已经确认的定时任务到期，都可能让同一 Agent 执行体系开始工作。L1 只回答“是谁或什么约定提出了一项工作”，不替后续模型决定研究计划，也不创建额外执行引擎。
保存关注主题并不自动触发无人值守监控。 已保存报告的定位和首页回读已有局部能力，订阅后持续生成、完整摘要和通知体验尚不能作为已交付产品来承诺。周报能生成，也不代表每一份周报都已经获准定时运行。
已确认的周期任务则有明确边界：中断后只在允许的宽限范围内接纳最近一期，不把所有错过的日期补成任务队列。是否完成研究、是否赶上约定时点、是否送出通知，分别读取原记录；时间过去本身不能补造成功。
非聊天入口与聊天入口最终复用原 Conversation/Task、Run、授权、Budget、Main/Child 和 Publication 机制。L1 的入口不同，并不产生第二套 Runtime。
三、L2 身份、会话与上下文：核对触发后准备任务
**账户身份先核验触发归属的真实主体、会话范围和可信变更来源；定时执行使用已确认的 Task/账户身份与原始授权记录，不能绕过认证。对话接纳与任务接续**再建立或复用消息与 Run，分开 Conversation 与单次执行身份。显式 Continue、普通追问、同消息重发以及已在运行的后台任务，不共享一种“继续”语义。
**Run 授权与能力准入**把来源、工具和能力范围绑定到原 Run。账户身份回答“是谁的任务”，Run 授权回答“这个 Run 具体可以做什么”，二者都在 L2 有自己的 Owner，任何一种入口都不得绕过。恢复后的同 Run 使用原获准事实，不让某次重新装配的上下文变成新的授权。
每次模型调用还需要准备材料。**上下文投影与拟合组装本轮历史、已获准工具结果、来源身份与容量限制；长期记忆提供受控的偏好或回忆；Main 上下文压缩决定哪些已结算历史可生成摘要、采纳并恢复；行动提醒协调**把资源或上下文压力转成不具有授权效果的建议。四者会相遇，但负责的事实不同。
跨 Run 会话历史压缩为新 Run 提供有来源的有损问答材料，保留用户原文；摘要必须对得回原问答，不能创造新的证据或权限。它与同 Run 的 Main 压缩不同，详见A41 独立机制文。
用户明确引用一份已提交研究时，新 Run 还可以绑定那份正文的确切版本及获准来源，而不是从整段聊天猜“上一份报告”。引用不可读或不合格，只省略对应材料，不阻断独立合法回答；旧研究也不会因此重新变成当前任务的执行状态。
方法也在这里装配：**Skill 方法与资源提供版本化研究方法，插件准入与生命周期约束扩展包的激活和冻结版本，生命周期事件与 Hook 执行**随后在相应生命周期点执行局部回调。Skill/Plugin 不是模型凭名称获得访问权限的捷径。
**个人专家与组合**也在这里准备。用户可保存个人专家的方法与偏好；新 Run 固定当时采用的版本，Main 先看到简短目录，选中的 Child 再接收对应方法。偏好的工具和 Skill 只是建议，不增加权限；显式预加载仍受资源读取和上下文容量约束。
L2 不只发生一次。进入 Research 的下一次 ModelTurn，仍可能重新投影上下文、读取当前合格结果或触发压缩，但不会因此重新创建 Run 或放大原授权。
四、L3 Agent 决策与运行：先分支，再多轮执行
这一层是模型真正参加决策的地方。**研究决策与综合**消费原问题与可见上下文，提出 Direct、Research 或 Coord；但可用分支由当前合法会话状态和资源条件决定，不能假定三种动作永远同时出现。
4.1 Direct：调用模型，但跳过研究工具
“行业集中度是什么意思？”可以直接形成短回答。它仍要走**模型目录与路由配置、模型调用与请求配置（Gateway）**和必要的 Provider 容量准入，但可以不执行 Web、PDF、Python 或金融工具。已有生产契约明确：Direct 内容也进入既有 Publication Owner，而非绕开正式发布机制。
模型选择需要看四个不同事实：客户端只能提出已配置的模型 ID；模型目录决定哪些路由与能力声明有效；本次 Gateway 绑定并发送实际配置；容量与 Run Budget 分别负责本次 Provider 资源和原 Run 的消耗。Deep 由 Run 资源管理批准，不等于切换模型；思考强度属于具体模型请求配置，也不自动证明供应商实际采用了更深思考。
4.2 Research：Main 与 Subagent 是执行主角
用户要求结合真实持仓与基金报告时，Main 在研究循环中选择下一步：取数、读资料、复算、要求 Child 探索、合成结论，或提出结束。模型决定研究方向，原运行时负责执行和结算；不需要为此再造一套 Supervisor 或通用 Workflow。
**Child 执行生命周期**不是入口的小功能。Subagent 拥有自己的局部执行、重试、取消、结果身份和 Handoff；但继续共享 Root Run 的授权、资源与可靠恢复事实，不能因为独立执行就建立第二套投资账本或公共发布权。Main 消费 Child 的真实结果，负责综合和最终面向用户的正文。
Web/PDF/Python 类型的 Child 和投资视角 Child 可以共用这套执行机制；是否值得委派由 Main 按任务判断。Child 读过某份原文、Main 只看过它提交的摘要，代表不同的证据消费事实，不可混用。
Child 也不是只能照着最初任务执行。发现假设不成立时，它可以在阶段结果中修订后续任务、方法提示和协作关系说明，下一轮沿新描述继续。已发送请求、用户原文、权限和预算不会被改写；新增材料仍受原来源检查。关系说明也不会自动赋予它读取或控制另一个 Child 的能力。
放大图：研究为什么是一个回环，而不是从 L1 到 L6 的单向流水线
主循环是 Main 决策 → Host 执行 → 工具回执／研究结果 → 新一轮 Main 决策。Main 可以在任何合规的研究时点委派 Child；Child 只提交本地结果和引用，Main 再决定是否补查或综合。一次循环结束不代表整个 Run 已完成，而在循环尚未结束时，已经成立的部分正文也可以经 Publication 独立提交。
图底部三条路径并不顺序执行：Publication 管内容版本，Recovery/HIO 区分合法等待唤醒与中断结算，观测与评估用独立证据解释错误。它们不能因为触及同一个 Run 就互相签发权限或重写事实。Direct 仍要调用模型 Gateway 和发布系统，只跳过研究工具。
4.2.1 研究如何继续、改道、等待与结束
上图放大的是一次模型回合的决策与观察。如果从完整任务看，还可以区分三个时间尺度；它们是同一架构的阅读视角，不是三套嵌套的调度器。
例如，用户询问“我的基金行业集中度是否变化”。检索命中一篇今天重新转载的基金报告，其持仓截止日可能在数月前。Main 应区分转载日、披露日和数据 as-of，再决定补查原文、使用标明时点的历史数据，或说明无法确认变化。搜到了“最新文章”，不等于拿到了“最新持仓”。
这不是完全缺少时间基础能力。当前 A36 时间语义已区分当前 UTC、不可变任务锚点和用户本地日期；时区按本次请求、账户偏好、设备提示依次取有效值，无可信来源就保留未知。账户保存和重启后的原请求重放已有真实持久化验证。美股日历也已有针对 2026 年 Nasdaq 常规时段的官方年历读取和有界解释：可以区分常规开闭市、休市和提前收盘，但年历不证明市场实际开盘，也不覆盖临时紧急变更，更不代表行情数据已经更新。普通真实模型研究实际消费日历的端到端验收仍未完成，不能把这套机制写成全自动时间判断已经可靠。
结束也是两个判断：Main 负责回答是否足够、需要哪些反证及不确定性解释；CompletionGate/Publication 处理合法的完成、资源耗尽、部分正文及实际发布。Run 终态不认证金融分析质量，resolved 也不等于任务已完成。Child、Reviewer 和多空视角可以在需要时参与，但不是每项研究都必经的外围循环。
研究循环怎样让观察改变下一步，另见投研 Agent 的关键不是多调用工具：Model、Harness 与 Loop。
4.3 Coord：协调原 Run，不是第三套 Agent
当同一个 Conversation 已有后台 Research，用户说“给刚才那项研究补上汇率因素”，普通新消息可以走 Coord。它必须关联合法的目标 Run，提出 inspect、correct 或 stop 等有界动作，再由原 持久控制事件、Correction、Continue 或终结 Owner 接纳真正的效果。模糊目标只澄清，不靠最新消息猜对象。
当前默认容量配置下，没有活动后台 Research 时普通 Entry 可选 Direct/Research；有活动后台 Research 时普通 Entry 可选 Direct/Coord。这只是经准入的动作集合，而不是先按 hasBackground 做一轮固定分类。显式 Research/Deep 可跳过普通分类，但仍受后台容量、Scope 总额、Provider 租约和原预算约束。
Coord 消耗短交互资格；它可以独立回答问题，也可能请求调整原研究。一个 Coord 消息接受了，不等于原 Research 已完成纠偏。 只有目标 Owner 的实际回执能说明效果。
五、L4 工具与领域能力：把模型的“去查”变成可核对的回执
模型提出一次工具动作，并不说明外部执行已经发生。**模型能力呈现与工具表面确认模型真正看到的可用工具与 Host 当次获准执行面一致；受控工具调用核对参数、执行、结算并留存回执；MCP 服务接入**负责远端连接、目录与协议通信。它们的权限、发送与结果职责不能互相替代。
读研究资料时，可复用 Web 研究读取、PDF 文档读取、知识检索、研究资料资产、隔离分析。基金相关的数据与计算则通过“金融工具”族提供：市场证据采集、投资事实读取、确定性投资计算、投资账本与确认、原子报告生产。财务或时间敏感任务还会使用时间语义与市场 Session 解释，必要时使用精确数值运算。
因此，不应把“证据与计算”当作每次请求的必经关卡。它们是L4 里可选、可组合的研究与领域工具能力。普通 Direct 完全可以跳过；复杂投研可以反复取数、重读和计算。
但有两条责任不能隐藏：第一，**研究结果与引用管理保留 ResultRef、版本和来源/派生谱系，不能因为模型改写摘要就把新断言变成原始事实；第二，投资账本与确认**的真实写入仍由业务 Owner 和用户确认要求控制，不能把模型建议当成已获批准的交易效果。
5.1 四个容易混淆的交接
工具链中经常出现“上一环节成功，却被当作下一环节也成功”的误判。以基金资料研究为例：
这里还有一个细小但重要的扩展约束：调用后 Hook 允许收窄模型可见投影，但不能把保留下来的原始数值、来源和路径悄悄换成另一套事实。Hook 的局部失败也不自动拥有整个 Run 的终结权。
当数据时点、覆盖率或金融口径不足时，正确结果可以是 Partial 或 Unknown。模型负责分析和解释缺口，确定性金融计算负责领域公式，Numeric 只执行合同规定的数值运算；三者并不互相认证。
六、L5 结果发布与交付：生成了内容，还不代表已经交给用户
**正文发布与内容准备负责候选正文、预览、版本修订、提交及其回执。Direct、Research 的部分结果与最终结果都要尊重它的内容处置权。私有产物保存、访问与分享另管 HTML、图表等 Artifact 的字节身份和下载许可；会话事件与历史交付**负责 SSE、顺序、History 重放和后台通知。
同一段文字可能经历不同事实：模型生成了正文、Publication 接纳了预览、数据库提交了修订、服务端推送了事件、客户端最终绘制了它。它们不是一个 completed。合法的部分正文可以先发布；后续工具失败不能把此前已提交的正文自动清空。
后台研究尤其需要这些区分。Research 可以不占住当前聊天输入，在另一个 Run 中持续执行、阶段性发布、到约定时点报告真实进度，最后提交 Final。at 和 by 的时间承诺、Research 完成时刻与消息通知时刻是不同事实；通知不会为了重发而重新跑一次研究。
6.1 一次“继续”，要先分清哪种情况
这也解释了为什么 Run 结束、正文已提交、用户被通知、浏览器已经显示 必须分成不同事实。观测应回到它们各自的真实回执，而不是依赖一条 completed 或最后一个 Trace Span。
七、L6 用户界面：交付对象必须能对回原 Run
**客户端交互与阅读**负责输入、消息时间线、Markdown 和引用卡片、研究面板、SSE 消费、历史重放，以及中途插话后的多个 Run 展示。
它不应该靠数组最后一条消息或一个全局 busy 决定所有消息归属。后台 Research A 尚未完成时，用户可以发起 Direct B；B 可以先完成，而 A 的阶段性结果和 Final 仍显示在 A 的研究视图下。History 快照也必须按照 Run 与 Publication revision 做有序还原，不把旧内容覆盖为新结果。
浏览器显示 answer-preview 不等于内容已经耐久保存；服务端说“事件已发送”也不等于浏览器完成绘制，更不等于用户真正阅读。L5 负责可交付的事实，L6 负责客户端实际呈现，两者分开是必要的。
界面也能把“选了谁”和“谁真的参与”分开：用户可搜索内置与个人专家，精确选择一位，或应用可编辑的多专家组合建议。协作区展示真正参与的 Child、公开 Peer 交流和受管 Reviewer 贡献，刷新后仍对应原 Run 与专家版本。选择不等于已调用，组合不是强制全员执行的工作流。
并非所有阅读都要重新研究。获准的组合首页可直接显示原报告的可配置组合总额、资产分组、估算与数据缺口，以及同份报告中的已确认策略版本；打开页面不会启动模型或投资写入。报告时点和未知状态必须保留，页面不能把旧快照包装成实时账户，也不能把策略展示说成新的投资批准。
八、用三次用户输入，走一遍交互分支
用户先说：“分析我的基金组合行业集中度，给出来源。”并显式开启后台研究。L1 形成任务输入，L2 认证并建立 Research A 和合法 grant；L3 Main 决定研究策略，可能委派 PDF Child 核对报告或 Python Child 做数据转换；L4 返回工具真实回执与来源；L5 可以先发布部分材料，再发布最终报告；L6 用 A 的身份显示研究卡片。
随后用户问：“行业集中度为什么重要？”新的短交互 B 可以由 L3 走 Direct，不等待 A 结束，也不读取 A 未获准公开的内部状态；短回答经过 Publication/History 到前端。
再随后用户说：“给刚才的 A 加上汇率风险。”新的短交互 C 经 L3 判断为 Coord，仅向明确的目标 Run 提出纠偏；原控制 Owner 接纳后，A 在合法安全点消费它，继续已有预算和恢复身份。C 不自己生成第二份金融分析账本。若 A 此时已结束，Coord 不能装作成功注入，用户可以再选择显式 Continue 建立关联的新 Run。
这套后台研究、短问、原目标纠偏与分开回读的路径，已有包含真实模型、持久 History 和浏览器的限定场景验证；但并发会话配置仍默认关闭。默认短交互/后台容量为 3/1，同时受 Scope 总额、交互实例容量和 Provider 容量约束。实例限制不是集群总限额，独立定时 worker 与控制唤醒另有各自的容量约束；这些局部验证也不等于生产部署或金融分析质量已经通过。
九、纵向能力柱为什么允许跨层，甚至断裂
六层是“什么时候需要完成什么交接”，Owner 是“谁对这项事实负责”。两种视角不是强制一一对应。
账户身份与 Run 授权与能力准入主要在 L2，但 L4 的 Host 执行和 L5 的来源相关处置都要消费原授权。长期记忆、上下文投影、Main 上下文压缩主要归 L2，却会在 L3 的下一次模型请求中再次出现。模型目录与路由配置、**模型调用与请求配置**主要服务 L3，但模型选择从 L6 发起并经 L1/L2 验证或绑定。
Child 执行生命周期在 L3 形成核心协作，在 L4 触达真实工具；研究结果与引用管理贯穿工具结果与发布消费。发布与交付主要在 L5，用户界面通过 L6 消费其合法版本。它们可以在图中形成连续或断裂的纵向柱子，不要求调用总是沿着柱子上下执行。
跨层显示不能制造新的权限：一根柱子覆盖几层，仍然只有原来的 Owner 能接纳、拒绝和结算自己负责的效果。纵向聚合只是为了让读者先理解系统，而不是新设一个父级模块。
十、Recovery、治理、观测为什么没有第七、第八层
Recovery 首先保住事实，不承诺复活任意执行。 浏览器重连只等待或重放已有结果；普通 Root 失去原执行者时，系统核验租约与 fence，按已保存正文结算为中断后的部分结果或失败，不在新进程自动重建旧执行图。已提交回执、用量和未知外部效果继续保留。只有经原控制通道接纳的人机/Child 唤醒，才使用保留下来的有限当前格式状态；用户显式继续则建立关联的新 Run，复用合格材料而不是旧执行权限。各项效果仍由原控制、资源、授权、Child、发布和历史 Owner 负责。
Authorization、Run Budget、Provider Capacity、内容信任、HIO分别在多个环节提供约束或交接，而不是集中到一个“安全层”。HIO 已实现请求、认证回复及向原消费者的交接，生产 Human Park 仍默认关闭；“模型发问—用户答复”也不自动等于来源获准或业务效果已批准。
研究调用证据留存与回读、Langfuse 和质量评估属于跨层观测与离线反馈。Trace 可以帮助定位哪次模型请求没看到工具结果，评估可以复核真实发布的答案是否符合原题；它们不承担发布、授权、终结或业务真值的决定权。请求成功、工具返回、正文提交和金融结论正确，也不应合并为一个成功率。
十一、这张图能回答什么，不能回答什么
这张六层图能在一分钟内告诉读者：用户问题进入哪几个阶段；Direct 为什么能跳过研究工具；Main 与 Subagent 为什么是核心；模型选择、Gateway、Memory、压缩、授权、工具、发布、前端分别出现在哪里；后台研究为何可以持续运行而前台继续回答。
它不能代替完整 Owner 图谱、源码调用关系、部署拓扑、精确的执行时序或线上质量报告。为保持可读性，主图将相近能力放进纵向能力柱，将五项 B 类金融模块与研究工具放在同一列，同时保留每个模块自己的实线标签；完整名字与编号由权威清单索引，实际授权与事实不因视觉聚合改变。
设计和排查时，先问三件事：模型此刻拿到了什么，哪个 Owner 真实接纳了效果，用户能回读哪份结果。 这比沿六层各找一个“成功”更有用：工具成功不证明研究充分，进程重启不授予旧执行权，正文生成也不等于交付完成。
附录：能力 Owner 与六层的主归属索引 {#capability-owner-links}
正文与主图使用模块名称；编号便于定位 模块详情。下表的“主层”不表示只在该层运行。
L1 · 任务输入与触发
L2 · 身份、会话与上下文
L3 · Agent 决策与运行
L4 · 工具与领域能力
L5 · 结果发布与交付
L6 · 用户界面
工程导航 E01–E03 不计入上述 47 个生产能力 Owner。各模块是否已实施、限定验证、生产启用，仍以自身 Spec 与验收记录为准。
贯穿全景的两条判断： 模型负责语义、系统负责执行事实用来区分模型的研究提议与 Host 的授权和回执；用户实际收到的结果才是任务终点要求把模型生成、正文提交和浏览器可见交付分别核验。六层是读懂交接的地图，而不是任何一层的自我成功声明。
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 value of the six layers is not to put the model into a fixed stream line, but to sort out each transition: who offers the job, who judges next step, who really implements it, and which results have been delivered to the user. This paper follows a request to explain what is being achieved. Six layers are map reading, not six processes; cross-layer participation does not change the responsibility of the modules for permission, facts and effects.
One user question, why six links?
The following is a teaching scene that runs through the main chain: users enter “Analysis of my fund portfolio to see industry concentration and exchange rate risk” in a chat box. It is used to explain the call relationship and is not an investment conclusion or real performance.
From the interface, there was only one thing that happened: to send a message and get an answer. Within the system, at least six different issues need to be addressed: what type of input or event triggers the work; which user and Run the trigger is attributed to and which context is permitted; how models are prepared to deal with; whether tools need to be called; which elements are actually approved for publish; and how the user interface is shown or restored.
Read this diagram as if you were tracking " Enter URL to Page Render " : advance along the request and then expand where there are branches and loops. Common conceptual questions can be answered directly, and when information is not available, they enter the study; while back-office research is still in operation, new information may be additional. These cases share a lower-level capability but are not the same fixed path.
47 modules: 60 seconds to read six floors
Imagine a line from my colleague: "Help me compare the two fund reports, tell me the difference, and keep it." From the time he clicks to receive the report, he can remember only six things:
If the second report is not read, the assistant may submit the confirmed portion, but the missing comparison cannot be described as completed. Conversely, keeping the report does not mean that it must have been displayed on the page.
The main figure below is the technical index, not the pre-reading examination. First reading can be done by skipping 47 module names, looking down only to the six questions; then the corresponding module should be opened when a step is required. The six floors are not a fixed approval process that must be completed for each request.
Interactive Navigation: A panorama in the original technology station supports clicking on the Owner label, full screen and zoom; the corresponding module is also available through the Owner Details.
Opens a condensable SVG panorama: 9 capability groups, 47 stand-alone module labels; research tools are combined with financial instruments, identity and authorization.
The six crossbands in the figure correspond to the six request processing links; each coloured pillar aggregates a set of capabilities for reading. Each line label represents a stand-alone Owner module; a light-coloured dotted line box represents a processing action or a description of the use, and indicates in the box " Source: correspond to Owner". The dotted line heading at the top is only a reading group, not a single module. Vertical groupings can be cut across layers or cut apart, but do not result in additional authorization or control. Horizontal stratification and vertical capacity are not mutually exclusive classifications: an Owner can participate in multiple layers, but its factual attribution will not be replicated.
Here's the L6 not that the interface only appears at the end. Sending buttons and model selections also occurs in L6, but L1 does not just receive chat messages: confirmed timing Task can also trigger Research. L2 then performed a definitive verification of the identity, Run and permission that triggered the association. Six layers describe the main chain, which does not mean that Run is created every time.
ii. L1 Task Input and Trigger: Agent does not necessarily start with chat frames
L1 is understood to be the starting point of the task, not the HTTP authentication layer. Users who click " Send " are the most common entry points; user offers to generate weekly reports, to click " Research Today ", or to have confirmed the expiration of a scheduled task, may allow the same Agent implementation system to begin. L1 answers only “who or what agreed on a job” and does not determine the research plan for the successor model or create additional implementation engines.
Saving the subject of concern does not automatically trigger unmanned surveillance. The location of the saved report and the return reading of the home page already have local capabilities and the continuous generation, complete summary and notification experience after subscription are not yet committed as delivered products. The production of weekly newspapers does not mean that each weekly newspaper is authorized to run on a regular basis.
The established periodicity has clear boundaries: The break is followed by the acceptance of the most recent issue within the permissible grace limits, without all missed dates being added to the task queue. Whether or not to complete the study, to meet the agreed point of time, to send a notice and to read the original record separately; the time past cannot itself compensate for success.
The non-chat and chat portals eventually revert to the original Conversion/Task, Run, Authorization, Budget, Main/Child and Publicisation mechanisms. The L1 entrance is different and does not produce a second set of Runtume.
III. L2 Identity, session and context: Checking trigger to prepare task
**The account identity validates the true subject, the scope of the session and the credible source of the change that triggers the attribution; the use of identified Task/account identity and original authorization records is performed on a regular basis and cannot be bypassed. Dialogue admissions and mandate renewals** create or return messages to Run, separate from Conversion and single execution status. Visible Continue, regular queries, re-transmitting of messages and running backstage tasks do not share a “continue” syntax.
** Run Mandate and Capability Access** Tied sources, tools and capabilities to the original Run. The account identification answers "who's mission," and Run is authorized to answer "what this Run can do," both of which have their own Owner in L2, and no one can go around. Recovered Run uses the previously authorized facts to prevent context from being converted into a new mandate.
Material is also required for each model call. ** context projection and proposed assembly of current round history, approved tool results, source identity and capacity limitations; long-term memory provides controlled preferences or memories; Main context compression determines which settled history can generate summary, adopt and restore; Action Alert ** Coordination transforms resources or context pressure into recommendations that do not have the mandated effect. The four will meet, but the facts of responsibility are different.
Cross-Run session history compresses the new Run to provide a source-friendly question-and-answer material and to retain the user ' s original; summary must be able to get back to the original question-and-answer session and cannot create new evidence or permission. It is different from the Main compression in Run, as detailed in the A41 Independent Mechanism.
When the user expressly cites a submitted study, the new Run can also bind the exact version of the text and the approved source, rather than guessing from the whole chat “the previous report”. The reference is not readable or qualified, only the corresponding material is omitted and does not block an independent legal answer; nor will the old study lead to a return to the status of the current mission.
The methodology is also assembled here:** the Skill methodology and resource availability version study, the activated and frozen version of the Plugin Access and Life Cycle Constraint Extension, and the Life Cycle Event and Hook Implementation** are then returned to the Executive Board at the corresponding life cycle point. Skill/Plugin is not a shortcut for the model to access permission by name.
** Individual experts and configurations** are also prepared here. Users can save the methods and preferences of individual experts; the new Run fixes the version used at the time, Main sees the short directory first, the selected Child then receives the corresponding method. Optimal tools and Skill simply suggest that permission is not added; visible preloading is still subject to resource reading and context capacity.
L2 not only happens once. The next ModelTurn to enter Research is still likely to re project context, read the current eligible results or trigger compression, but will not recreate Run or zoom in the original authorization.
IV. L3 Agent decision-making and operation: first branch, then multiple rounds of implementation
This layer is where the models actually participate in decision-making. ** Research on decision-making and integration of ** consumption problems with visible context, suggesting Direct, Research or Coord; however, the available branch is determined by the current legal status of the session and resource conditions, and it cannot be assumed that the three actions will always occur simultaneously.
4.1 Direct: Call models but skip research tools
What does “industry concentration” mean?” A short response can be made directly. It continues to follow** model catalogue and route configuration, model call and request configuration (Gateway)** and necessary Provider capacity access, but can not execute Web, PDF, Python or financial instruments. There is a clear production contract: Direct content also goes into the existing Publicization Owner rather than bypassing the formal publish mechanism.
Model selection requires four different facts: the client can only present the configured model ID; the model catalogue determines which route and capability statements are valid; this Gateway binds and sends the actual configuration; and capacity is responsible for the consumption of this Provider resource and the original Run, respectively. Deep, approved by Run Resource Management, does not amount to a switching model; the intensity of reflection is part of a specific model request configuration and does not automatically prove that suppliers actually use deeper thinking.
4.2 Research: Main and Subagent are key players in implementation
When the user requests a combination of real hold and fund reports, Main selects next step in the study cycle: take numbers, read information, double-count, ask Child to explore, synthesize conclusions, or put an end to them. The model determines the direction of the study, which was operated with responsibility for implementation and settlement; a further set of Supervisor or Universal Workflow is not required for this purpose.
** Child execution life cycle** is not a small entry function. Subagent has its own partial execution, retry, cancellation, outcome identity and Handoff; however, the continuing sharing of Root Run ' s mandate, resources and reliable restoration of facts cannot lead to the creation of a second set of investment books or public publish rights as a result of independent execution. Main Consumer Child Real Results, which combines and ends up being user-oriented.
The Web/PDF/Python-type Child and Investment Perspectives Child could share this implementation mechanism; whether or not it would be worthwhile to delegate Main as a task judge. Child has read one of the original texts and Main has only read the summary submitted by it, which represents different evidence of consumption facts and cannot be used in combination.
Child did not just follow the original mission. When it is found that the assumption is not valid, it can revise the follow-up tasks, methodological tips and collaborative relationship statements in the outcome of the phase and continue with the next round along with the new description. Requests sent, user originals, permission and budgets will not be rewrited; additional material is still checked from the original source. The relationship statement does not automatically empower it to read or control another Child.
Magnifying: Studying why it's a loop, not a one-way waterline from L1 to L6
The main cycle is Main Decision-Making → Host executes tool validation/research results → new round of Main decision-making. Main can assign Child at any time of compliance study; Child only submits local results and references, and Main decides whether to re-check or synthesize. The end of a cycle does not mean that the entire Run has been completed, and part of the text that has been set up can also be submitted independently through Publicization before the end of the cycle.
The three paths at the bottom of the figure were not sequentially executed: the Public Management Content version, Recovery/HIO distinguished between lawful waiting for awakening and interruption of settlement, and observations and assessments interpreted errors using independent evidence. They can't issue permission or rewrite facts to each other because they touch the same Run. Direct still calls models Gateway and publish, skipping research tools.
4.2.1 Study how to continue, change, wait and finish
The above figure is amplified by a model round of decision-making and observation. It is possible to distinguish between three timescales when looking at a complete mission; they are the reading perspective of the same structure and not three nested dispatchers.
For example, the user asked “Does the concentration of my fund industry change?” The deadline for the retrieval of a copy of the Fund ' s report, reproduced today, may have been months ago. Main shall distinguish between the date of reproduction, the date of disclosure and the data as-of and then decide to recheck the original language, to use historical data indicating the point of time or to indicate that the change cannot be confirmed. The search for the “latest article” did not amount to a “latest hold”.
This is not a complete lack of time-base capacity. The current A36 time semantics distinguish between the current UTC, the non-changing task anchor and the user ' s local date; the time zone takes the valid value in order of this request, the account preference, the device hint and leaves unknown without a trusted source. The original request to reset after the account has been saved and restarted has been authentically and permanently validated. The U.S. stock calendar also has an official calendar reading and interpretation for the regular period of the Nasdaq in 2026: a distinction can be made between regular closed markets, closed markets and early closings, but the calendar does not prove that the market is actually open, nor does it cover temporary emergency changes, much less that the status data has been updated. The end-to-end acceptance check of the General Real Model study of the actual consumption calendar is still not complete and the mechanism cannot be written in full automatic time judgement.
The end is also a two-judgement: Main is responsible for responding to the question of whether it is sufficient, what is needed as a counter-argument and for the explanation of uncertainty; Completing Gate/Publication deals with legal completion, depletion of resources, partial text and actual publish. Run ultimately does not certify the quality of financial analysis, nor does resolved mean that the task has been completed. Child, Reviewer and the multi-space perspective can be involved if needed, but not in the outer circle through which every study has to go.
Studying how the cycle changes the observation of next step, see also that the key to research Agent is not multi-call tools: Model, Harness and Loop.
4.3 Coord: Harmonizing old Run, not third set Agent
When the same Conversion has backstage Research, the user says, "Add the exchange rate factor to the study just now," and the usual news can go Coord. It must relate to a legitimate target, Run, suggesting that there are bounds of action such as inspect, corp or stop, and then accept the real effect by the original Protracted Control Event, Corraction, Continue or terminate Owner. Fuzzy targets only clarify and are not subject to the latest information.
Under the current default capacity configuration, general Entry is available for Direct/Research when there is no active backstage Research; normal Entry is available for Direct/Coord when there is active backstage Research. It's just an access assembly of actions, not a round of regular categories by the HasBackground first. Visible Research/Deep can skip the general category, but is still subject to backstage capacity, the sum of Spope, the Provider lease and the original budget.
Coord consumes short interactive qualifications; it can answer questions independently or may request adjustments to the original study. A Coord message accepted is not the same as the original Research completed correction. Only the actual confirmation of the target Owner can show the effect.
L4 Tools and area capabilities: transforming the model's “go check” into a verifiable response
The model presents a tool action that does not indicate that external execution has taken place. ** Model capability presentation is consistent with the tool surface recognition model ' s real available tools that host was approved for implementation; controlled tool call checks parameters, executes, settles and retains back receipt; MCP Service Access** is responsible for remote connectivity, catalogues and protocol communications. There is no substitute for their permission, sending and outcome duties.
When reading research materials, Web research reading, PDF documentation reading, intellectual retrieval, research assets, isolation analysis can be used. The Fund's data and calculations are provided through the “financial instruments” community: market evidence collection, investment fact reading, firm investment calculations, investment books and confirmations, and production of atomic reports. Financial or time-sensitive tasks are also explained by time semantics and market session, using precise numerical calculations where necessary.
Therefore, “evidence and calculation” should not be considered a mandatory entry point for each request. These are the L4-optional, portfolioable research and area tools capabilities. Normal Direct is perfectly passable; complex research can be repeated, read and calculated.
However, there are two responsibilities that cannot be concealed: first,** the results of the study and the reference to the management reservation Result Ref, version and source/spectrum systems, and that the new assertion cannot be made original by the model ' s rewriting of summary; second, the true writing of the investment book and confirmation** remains controlled by the business Owner and user confirmation requirements, and the model proposal cannot be considered as having the effect of an approved transaction.
5.1 Four easily confused intersections
There are frequent miscalculations in the tool chain that “the first link succeeds, but is considered the next”. In the case of the Fund, for example:
There is also a small but important extension constraint: Hook, after call, allows narrow models to be projected, but the original values, sources and paths retained cannot be quietly replaced by another set of facts. Hook’s partial failure does not automatically have the end right for the entire Run.
The correct result can be either Partial or Unknown when data are time-bound, covered or of insufficient financial calibre. The model is responsible for analysing and explaining gaps, determining the formula for the area of responsibility for financial calculations, and Numeric only performs numerical calculations as stipulated in the contract; the three are not mutually authenticated.
VI. L5 Results publish and Delivery: Generating content does not mean that it has been handed over to users
** The main text publish and the content are responsible for the submission, preview, revision, submission and receipt of the text. Direct, Research, in part, and ultimately respect its content disposition. byte identification and download permission for the preservation, access and sharing of privately owned products, other HTMLs, charts, etc., Artifact; session event and historical delivery** responsible for SSE, sequence, History replay and backstage notification.
The same text may experience different facts: the model generates the body, the Publicization accepts previews, the database submits revisions, the service sends events, and the client ultimately draws it. They're not a completed. A legitimate part of the body can start with publish; failure of the follow-up tool cannot automatically empty the body previously submitted.
These distinctions are particularly needed in back-office research. Research can leave the current chat input unoccupied, keep running, phase publish in another Run, report real progress by the appointed point, and finally submit Final. At and by time commitment, Research completion is a different fact than the time of notification; the notification does not rerun the study for the reissue.
6.1 One “continuation” to distinguish which situation
This explains why Run ended, the text was submitted, the user was notified, the browser had shown that it had to be divided into different facts. The observations should revert to their respective true responses, rather than relying on a completed or last Trace Span.
VII. L6 User interface: delivery objects must be able to return to the original Run
** Client interactivity and reading** is responsible for input, message time lines, Markdown and quote cards, research panels, SSE consumption, historical replay, and multiple Run displays after mid-intersects.
It should not depend on the last message of the array or on a global busy to determine all the messages. Backstage Research A can start Direct B; B can do it first, while Stage A results and Final are still shown in A's research view. History snapshots must also be systematically reduced according to Run and Publicisation and do not cover old content into new results.
Browser displays answer-preview not as much as content has been saved for a long time; the service says that "incident sent" does not mean that the browser has finished drawing, much less that the user is actually reading. L5 Responsibility for deliverables, L6 responsibility for the physical appearance of the client, is necessary.
The interface can also separate “select whom” from “who is really involved”: users can search for in-house with individual experts, select one precisely or apply editable multi-expert group recommendations. The collaboration area presents truly participatory Child, open Peer communication and managed Reviewer contributions, updated to match the original Run and expert versions. The selection does not amount to a call and the combination is not a workflow that is mandatory for full staff.
Not all readings have to be re-examined. The first page of the approved combination shows the total configuration of the original report, the grouping of assets, estimates and data gaps, as well as the confirmed version of the strategy in the same report; opening the page does not trigger model or investment writing. The reporting time and state of the unknown must be kept, and the page cannot wrap up old snapshots into real-time accounts, nor can the strategy show be described as new investment approvals.
viii. Enter with three users, walk through interactive branches
The user said, "Analyss my portfolio industry concentration and gives us a source." And it opened the back-office study. L1 forms task input, L2 certifies and establishes Research A and legal Grant; L3 Main decides to study strategy, possibly assigning PDF Child to the reconciliation report or Python Child for data conversion; L4 returns the tool 's true confirmation and source; L5 can start with publish partial material and then publish final report; and L6 displays the research card as A.
Then users asked, "Why is industry concentration important?" A new short interactive B can be left by L3 go Direct, without waiting for A to end, or read A is not allowed to open an internal state; short answers pass Publicization/History to the front end.
And then the user said, "Put exchange rate risk on the A just made." The new short interactive C, judged by L3, is Coord, with only a correction to the clear target Run; after the original control over Owner was accepted, A consumes it at a legal security point and continues to have a budget and restore identity. C does not generate a second financial analysis book. If A is over at this time, Coord cannot pretend to be a successful injection, and the user can then select the new Run associated with Visible Continue.
This set of background studies, short questions, original target correction and separate rereading paths has been validated with a defined scenario containing real models, persistent History and browser; however, the side session configuration is still defaulted to close. The default short interactive/backstage capacity is 3/1 and is subject to a total of Scope, interactive case capacity and Provider capacity. Examples are not total cluster limits, and independent timing and control are subject to separate capacity constraints; these partial validations do not amount to production deployments or the quality of financial analysis passed.
IX. WHY THE LOTAL CAPACITY POWER allows for stratums and even fractures
The sixth floor is "when do you have to finish the handover?" The Owner is "who is responsible for this fact." The two perspectives are not mandatory.
The account identity and Run mandate and capacity access are mainly in L2, but both L4 Host execution and L5 source-related disposal require consumption of the original authorization. long-term memory, context Projection, Main Z48Z compression is mainly L2, but will appear again in the next model request for L3. Model catalogue and route configuration,** model call and request configuration** major service L3, but model selection was initiated from L6 and validated or bound by L1/L2.
Child implemented life cycle L3 to form a core collaboration to reach real tools in L4; research results and references to management cross-cutting tool results and publish consumption. publish and delivery are mainly in L5, with the user interface consuming its legitimate version through L6. They can form continuous or fractured vertical columns in the picture and do not require calls to always follow the columns up and down.
The cross layer shows that no new permission can be created: a pillar covers several layers, and only the original Owner can accept, reject and account for its own responsible effects. Vertical aggregation is intended to allow readers to understand the system first, rather than creating a new parent module.
X. RECUSH, GOVERNANCE, WHY NOT THE SEVENTH AND EIGHTH LEVELS
Recovery first keeps the facts and does not promise to revive arbitrary execution. Browser reconnecting only waits for or resets the results; when regular Root loses the original implementer, the system verifies the lease and the fence, either as part of the result or as a failure of the saved text settlement, and does not automatically recreate the old performance chart in the new process. Repository, usage and unknown external effects have been submitted and continue to be retained. The limited current format status retained was used only when the human machine/child was awakened through the original control channel; the user-in-chief continued to create the associated new Run, reusing the eligible material instead of the old permission. The effects remain the responsibility of original controls, resources, authorizations, Child, publish and History Owner.
Authorize, Run Budget, Provider Capability, Content Trust, HIO provide restraint or interface at multiple points rather than concentrating on a “security layer”. HIO has fulfilled the request, authentication response and handover to the original consumer, and the production of Human Park is still by default closed; nor is the "model inquiry-user response" automatically equal to the source approval or business performance approval.
Research calls for evidence retention and rereading, Langfuse and quality assessment are cross-layer observations and offline feedback. Trace can help locate which model requests did not see the results of the tool and assess whether the answers to the true publish can be reviewed for conformity with the originals; they do not have the authority to decide on publish, authorization, termination or real value of the business. Requests for success, tool returns, text submissions and financial conclusions are correct and should not be merged into a single success rate.
Eleven. What's the answer?
This six-layer map tells readers in one minute what stages the user question is going to; how Direct can skip the research tool; why Main and Subagent are the core; where model selection, Gateway, Memory, compression, authorization, tools, publish, front-end each appears; and why back-office research can continue and the front desk continues to answer.
It is not a substitute for complete Owner spectrograph, source call relationships, deployment scaling, precise execution time series or online quality reports. In order to maintain readability, the master chart places proximity capabilities in the vertical capacity column, places five B-type financial modules in the same column as research tools, while retaining each module ' s own line label; complete names and numbers are indexed by authoritative lists, and actual authorizations and facts are not altered by visual aggregation.
When designing and rehearsing, three things are asked: what the model has at this point, which Owner actually accepts, which results the user can read back. This is more useful than finding a “success” on each of the six floors: the success of the tool does not prove sufficient research, the process is restarted without the granting of old enforcement powers, and body creation is not equivalent to delivery completion.
Appendix: Capability Owner and the index of primary ownership of the six layers
The text and the master map use the name of the module; numbering allows the location of the module details. The “main layer” in the table below does not mean that only the layer is operated there.
L1 Task input and trigger
L2 Identity, Session and context
L3 Agent Decision Making and Operation
L4 Tools and area capacity
L5 Results publish and Delivery
L6 User interface
Project navigation E01-E03 excludes 47 production capacity above, Owner. The modules have been implemented, certified and made operational, still based on their own Spec and acceptance check records.
The model is responsible for semantics, the system is responsible for executing the authorization and endorsement of the research proposal used to distinguish the model from Host; the result actually received by the user is that the end point of the task requires that the model be generated, the text submission and the browser be delivered separately. The six floors are maps that understand the intersection, not self-success statements on any level.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.