本页目录

📘 第 11 课:What are managed agents?(什么是托管智能体)

官方课程名:What are managed agents?(Curriculum 第三段「Managed Agents」开篇) 官方地址:https://anthropic.skilljar.com/claude-platform-101 归档路径:Claudeexams/20-课程/课程003 the AI Platform 101/ 针对认证:CCAR-P(架构师 Professional) 归档日期:2026-10-14 ⚠️ 语料来源:官方 Skilljar 课程正文(用户提供,需登录)。CC BY-NC-SA 4.0,仅供个人学习。

⭐⭐ 本课权重提示:托管光谱的终点课 + 三大领域的场景汇。① 领域 1(17%):考纲原词「设计多智能体系统与编排策略」——本课例 3 的 coordinator + specialists 就是那张图的场景版;② 领域 4(16%):考纲要求「把评估做成验收标准并作为改动闸门」——本课的 rubric + 独立 grader 就是场景版;③ 领域 5(14%):permissions policy 拦住敏感动作等人类批准(HITL)。三课铺垫在此收口:第 7 课「自建 ↔ 托管」光谱、第 10 课「托管智能体默认自带 caching+compaction」。 🔗 一句话总纲(Recap 末句):You define what done looks like. the AI works until it gets there.(你定义「完成」长什么样,the AI 干到达标为止。)


🎯 本课学习目标

⚠️ 本课官方未提供 What you'll learn 区块。以下按文稿结构反推:

  1. 能定义 Managed Agents:构建与规模化部署 agent 的 API 套件,agent 循环由官方托管。
  2. ⭐⭐ 能列全 八大构件(Agents / Sessions / Environments / Tools / MCP / Memory / Outcomes / Multi-agent coordination)。
  3. ⭐⭐ 能讲清例 1 的评估闭环:rubric 定义「完成」→ 独立 grader 独立上下文评分 → 迭代到达标。
  4. ⭐⭐ 能讲清例 3 的编排模式:coordinator 委派 specialists + 共享文件系统 + 汇总合成。
  5. ⭐ 领域 5 落点:permissions policy 在敏感动作前拦下等人类批准。
  6. 能识别记忆的模式标注用法(「像两周前的 DNS TTL 事故」→ 带着上下文开工)。

📖 一、Managed Agents 是什么

官方原句: the AI Managed Agents is a suite of APIs for building and deploying agents at scale.* You define agents with specific tools, personas, and capabilities. You configure sandbox environments with the right packages and network controls. Then you fire off sessions from your own application, and the AI does the work inside an isolated container* with full file system access, bash execution, and web search.

精译:the AI Managed Agents 是一套用于构建与规模化部署智能体的 API。你定义 agent(配好工具、人格与能力),配置好沙箱环境(装对依赖包、设好网络管控),然后从你自己的应用里发出会话,the AI 就在隔离容器里干活 —— 完整文件系统、bash 执行、网页搜索一应俱全。

⭐⭐ 托管的还是那个循环(本课第一论断)

官方原句: Under the hood, this is an agent loop: the AI reasons, calls a tool, reads the result, and repeats until the job is done. If you've built agents before, you've probably written this kind of loop yourself. Managed agents takes that same loop and hosts it on the AI provider's infrastructure, so you don't have to run it.

精译:底层还是那个 agent loop:AI 推理、调工具、读结果、循环到达成为止。你要是写过 agent,这个循环八成自己亲手写过。Managed Agents 做的事就是把这同一个循环托管到官方基础设施上——你不用自己跑。

🔗 「托管」光谱闭环:第 7 课 server tools「平台执行」(光谱第一档)→ 本课「整个循环托管」(终点档)。第 4 课那句 You own the loop and the tools 在托管模式下由官方代持——但责任划分的设计逻辑不变,这正是判断题的出题点。


📖 二、例 1:会干活的 Kanban 板(Kanban board that does the work)

官方原句: Picture a Kanban board sitting on top of managed agents. You drag a ticket into the "in progress" column, and that fires off a session automatically.**

精译:想象一块架在 managed agents 之上的看板。把工票拖进「进行中」栏,一个会话就自动发出。工票写着「优化网站性能」,接下来发生三件事:

步 官方叙事 构件
1 Your back end creates a session. Sessions(你的后端创建会话)
2 The session points to an environment you configured with Lighthouse and Puppeteer pre-installed. Environments(预装好审计工具的环境)
3 Your GitHub repo gets mounted into the container. 隔离容器(代码库挂载进来)

⭐⭐ 评估闭环(领域 4 的场景版)

官方原句: Now the AI has the codebase, the tools, and a rubric that defines what done looks like: Lighthouse score above 90 / No render-blocking resources / All images lazy loaded.

精译:此刻 AI 手上有代码库、工具,还有一份定义了「完成长什么样」的评分量表:Lighthouse 90 分以上、无渲染阻塞资源、图片全部懒加载。

官方原句: the AI runs the audit, then starts compressing images, inlining CSS, and deferring scripts. Every tool call streams back to the board in real time through the event stream.

精译:AI 跑一遍审计,开始压图、内联 CSS、延后脚本。每次工具调用都经事件流实时回传到看板——你能看着它干活。

官方原句: Then the rubric kicks in. A separate grader, running in its own context window, evaluates the output against your criteria. the AI reads that feedback, goes back in, fixes what it missed, and resubmits. In the demo, that loop takes the Lighthouse score up to 96.

精译:然后量表生效。一个独立的评分器,跑在它自己的上下文窗口里,拿你的标准去评产出。AI 读了反馈,回去补漏,再提交。演示里这一轮循环把 Lighthouse 分推到了 96。

⭐⭐ 考点拆解:① rubric(人定的验收标准)+ grader(独立上下文的评分器)=考纲「评估作为验收标准与改动闸门」的场景实现;② 「独立上下文」防的是运动员兼裁判——同一窗口自评会漂;③ 迭代闭环:feedback → fix → resubmit 直到达标。

官方原句: You can drag a second ticket over while the first is still running. Two sessions, two containers, two separate tasks running in parallel.

精译:第一个会话还在跑,你就能拖第二张工票。两个会话、两个容器、两件事并行。


📖 三、例 2:带记忆的定期研究 agent(recurring research agent)

官方原句: Here's a different shape of agent: one whose job is to track prices and plan changes across every SaaS tool your company pays for, with a report ready before stand-up.

精译:另一种形状的 agent:盯你公司付费的每一款 SaaS 工具的价格与套餐变动,报告在晨会前备好。

每轮运行四步:

# 官方原句 构件
1 Searches the web for current pricing pages, checks for plan tier changes, and flags new features web search(第 7 课 server tool)
2 Runs a cost analysis in Python inside the sandbox code execution(沙箱内算成本)
3 Uses an Excel spreadsheet skill and writes an executive summary Skills(第 8 课:规程与模板)
4 Posts a link to Slack and creates a review task in Asana, both through MCP servers** MCP(第 9 课:第三方连接)

官方原句: The agent also reads from and writes to a memory store. Before it starts, it checks what it found last week. After it finishes, it stores what changed. So next Monday's report can say "compute costs are 15% lower since last week" instead of listing the same static pricing data every time.

精译:agent 还会读写记忆库:开工前先查上周发现了什么,收工后存下这周变了什么。于是下周一的报告能写出「计算成本较上周降了 15%」,而不是每次都列同一份静态报价。

⭐⭐ 考点:这是第 10 课 memory tool 的增量用法——记忆不是存储数据本身,而是存储差量与结论(delta + conclusion),让报告从「快照」升级为「趋势」。判断题爱问「记忆该存什么」。


📖 四、例 3:多智能体事件响应(incident response)

官方原句: Now imagine an alert fires from your monitoring stack. A custom tool on your back end receives the alert payload and sends it into a new session as a tool result. This session uses multi-agent coordination.

精译:想象监控体系里警报触发。你后端的自定义工具收下警报载荷,作为一条工具结果送进新会话。这个会话用的是多智能体协同。

⭐⭐ 编排三结构(领域 1 考纲原词的场景版)

官方原句: - A coordinator agent receives the alert and delegates to three specialists. - **Each specialist runs in its own context window on the same shared file system. - The specialists report back, and the coordinator synthesizes their findings into a single incident summary.**

精译:① 协调者 agent 接下警报,委派给三个专家。② 每个专家跑在各自的上下文窗口里,共享同一个文件系统。③ 专家回报,协调者把发现合成为一份事故总结。

⭐⭐ 考点:这就是考纲领域 1「设计多智能体系统与编排策略」的官方场景版。三个特征连背:委派(delegates)+ 隔离上下文、共享文件(own context windows + shared file system)+ 汇总合成(synthesizes)。第 6 课 subagents 思想(专精子任务、主对话保干净)在此落地成产品。

⭐⭐ 权限策略与人类批准(领域 5 落点)

官方原句: Before the summary goes to Slack, the permissions policy fires. You see the draft on screen, approve it, and the message goes out. Sensitive actions wait for a human.**

精译:总结发去 Slack 之前,权限策略先触发。你在屏幕上看到草稿,批准后消息才出去。敏感动作等人点头。

🔗 呼应第 4 课 HITL 分界与第 9 课工具过滤:托管不等于放权——发布类敏感动作仍由策略闸门 + 人类批准把关。

记忆的模式标注(memory 的第三种用法)

官方原句: The coordinator checks past incidents in the memory store and flags a pattern: "this looks like the DNS resolution issue from two weeks ago that was caused by a misconfigured TTL." The next time a similar alert fires, the agent starts with that context instead of diagnosing from scratch.

精译:协调者查记忆库里的历史事故并标注模式:「这像是两周前那次 DNS 解析故障——根因是 TTL 配错」。下次类似警报响起,agent 带着这条上下文开工,而不是从零诊断。

⭐ 三例记忆用法对照:例 1 无记忆(单任务达标即止);例 2 记差量(周报趋势);例 3 记模式(历史相似事故 → 秒级定位)。


📖 五、八大构件(The building blocks · 必背清单)

官方原句: managed agents gives developers the tools to deliver a fully managed, stateful agent experience built on:

构件 官方定义 对应前课
Agents definitions with specific tools, personas, and capabilities 定义三件套:工具+人格+能力
Sessions individual runs you fire off from your own application 你的应用发起的单次运行
Environments sandboxes with the right packages and network controls 第 7 课 sandbox 的产品化
Tools including custom tools on your back end 第 4–5 课
MCP connections to services like Slack and Asana 第 9 课
Memory a store the agent reads before starting and writes to when done 第 10 课 memory tool
Outcomes rubrics and graders that define and check what done looks like 领域 4 评估闸门
Multi-agent coordination coordinators delegating to specialists 领域 1 编排

⭐⭐ 考点:「定义(Agents)→ 运行(Sessions×Environments)→ 连接(Tools/MCP)→ 记忆(Memory)→ 验收(Outcomes)→ 编排(Multi-agent)」六段结构记忆;考题把任一构件配错定义即为干扰。


📌 六、Recap(官方六条要点)

# 官方原句 一句话
1 the AI Managed Agents is a suite of APIs for building and deploying agents at scale, hosted on the AI provider's infrastructure. 规模化构建部署的 API 套件,官方托管
2 It runs the familiar agent loop — reason, call a tool, read the result, repeat — inside an isolated container with file system access, bash execution, and web search. 还是那个循环,装进隔离容器
3 Sessions run in environments you configure, work in parallel, and stream tool calls back to your app in real time. 会话并行、事件流实时回传
4 Rubrics and separate graders let you define success criteria; the AI iterates until it meets them. 量表+独立评分器,迭代到达标
5 Memory, MCP servers, custom tools, permissions policies, and multi-agent coordination round out the stateful agent experience. 记忆/MCP/自定义工具/权限策略/多智能体凑齐有状态体验
6 You define what done looks like. the AI works until it gets there. 你定义完成,AI 干到达标

🧠 本课 → CCAR-P 考点映射

本课内容 对应领域 权重
多智能体编排(coordinator+specialists+共享文件系统) 领域 1 Solution Design & Architecture 17%
rubric + 独立 grader = 评估闸门场景版 领域 4 Evaluation, Testing & Optimization 16%
托管循环、并行会话、事件流、容器隔离 领域 3 Integration 19%
permissions policy + 敏感动作人审(HITL) 领域 5 Governance, Safety & Risk 14%
托管智能体默认 caching+compaction(第 10 课钩子收口) 领域 2 13%
「你定义完成」的开发者叙事 领域 7 Developer Productivity 7%

⭐ 战略结论:本课是五个领域的场景交汇点,且两处直接命中考纲原句(多智能体编排、评估闸门)。三例各带一个考点模块(评估闭环 / 记忆差量 / 编排+人审),出题密度全课最高之一。


🔤 本课新生词(已回灌 30-单词表/03,2026-10-14)

词汇 音标 释义 备注
⭐⭐ Managed Agents — 托管智能体(产品专名) 托管循环的 API 套件 → 精讲 047
⭐⭐ session / environment /ˈseʃn/;/ɪnˈvaɪrənmənt/ 会话/环境 八大构件之二三 → 精讲 048;session 第 10 课已收,此处升级
⭐⭐ rubric / grader /ˈruːbrɪk/;/ˈɡreɪdər/ 评分量表/评分器 定义并检查「完成」→ 精讲 049;领域 4 闸门
⭐⭐ coordinator / delegate(v.) /koʊˈɔːrdɪneɪtər/ 协调者/委派 多智能体编排 → 精讲 050;表 01-062、02-003 delegate 已收
⭐⭐ persona /pərˈsoʊnə/ 人格、角色设定 agent 定义三件套之一 → 精讲 051
⭐⭐ synthesize /ˈsɪnθəsaɪz/ 合成、汇总 coordinator 合成发现 → 精讲 052
⭐ stateful /ˈsteɪtfl/ 有状态的 stateful agent experience;精讲 046 stateless 的正对照
⭐ event stream — 事件流 streams back in real time through the event stream
⭐ mount(v.) /maʊnt/ 挂载 GitHub repo gets mounted into the container
⭐ fire off — 发出、触发 fire off a session;an alert fires 同族
⭐ kick in — 生效、启动 Then the rubric kicks in
⭐ at scale — 规模化地 building and deploying agents at scale;对照第 1 课 scale(v.)
⭐ permissions policy — 权限策略 the permissions policy fires(领域 5 闸门)
⭐ isolated /ˈaɪsəleɪtɪd/ 隔离的 isolated container;表 01-014 已收
⭐ specialist /ˈspeʃəlɪst/ 专家(智能体) coordinator 委派对象
⭐ memory store — 记忆库 例 2/3 用法;对照精讲 045 memory tool
suite /swiːt/ 套件 a suite of APIs(⚠️ 不读 suit)
Lighthouse — Lighthouse(网页审计工具) 例 1 环境预装;分数 90→96
Puppeteer — Puppeteer(浏览器自动化工具) 例 1 环境预装
Kanban board — 看板 例 1 场景载体
render-blocking — 渲染阻塞的 No render-blocking resources(rubric 项)
lazy loaded — 懒加载的 All images lazy loaded(rubric 项)
inline(v.) /ɪnˈlaɪn/ 内联 inlining CSS
defer(v.) /dɪˈfɜːr/ 延后 deferring scripts
resubmit /ˌriːsəbˈmɪt/ 重新提交 fixes what it missed, and resubmits
recurring /rɪˈkɜːrɪŋ/ 定期循环的 a recurring research agent
stand-up — 晨会 a report ready before stand-up
plan tier — 套餐档位 plan tier changes
cost analysis — 成本分析 Runs a cost analysis in Python
spreadsheet /ˈspredʃiːt/ 电子表格 an Excel spreadsheet skill
executive summary — 高管摘要 writes an executive summary
static /ˈstætɪk/ 静态的 the same static pricing data(对比差量记忆)
alert /əˈlɜːrt/ 警报 an alert fires
monitoring stack — 监控体系 from your monitoring stack
payload /ˈpæloʊd/ 载荷 the alert payload
incident summary — 事故总结 a single incident summary
misconfigured /ˌmɪsˈkɒnfɪɡərd/ 配置错误的 a misconfigured TTL;mis- 族(misfire 016 同缀)
TTL — 存活时间(DNS 术语) caused by a misconfigured TTL
from scratch — 从零开始 instead of diagnosing from scratch;表 01-004 scratch 已收
pattern(n.) /ˈpætərn/ 模式、规律 flags a pattern(记忆的第三种用法:模式标注)
shared file system — 共享文件系统 specialists 隔离上下文但共享文件
host(v.) /hoʊst/ 托管 第 7 课已收,此处复现(hosts it on the infrastructure)
audit /ˈɔːdɪt/ 审计 表 02-058 已收,此处复现(runs the audit)
bash execution — bash 执行 第 7 课已收,此处复现(容器三件套之二)
web search — 网页搜索 第 7 课已收,此处复现(容器三件套之三)
parallel /ˈpærəlel/ 并行的 two separate tasks running in parallel

✅ 本课进度

  • [x] 全文精译与架构师级拆解(2026-10-14)
  • [x] ⭐⭐ 记录八大构件、三例各自的考点模块(评估闭环/记忆差量与模式/编排+人审)
  • [x] 记录「托管光谱闭环」(第 7 课 → 本课)与「You define what done looks like」总纲
  • [x] 生词回灌 30-单词表/03-考试核心生词与术语表.md(2026-10-14:新增速查 41 条 → 全表 468 条(与表内重复 5 词追加「第 11 课复现」);新升级精讲 6 条 047–052)
  • [x] 保存本课英文原文(26-English-11-What-are-managed-agents.md)
  • [ ] 跑通本课 runnable demo(需 Console 账号 + Managed Agents 权限)
  • [ ] 官方未提供 What you'll learn 与练习区块 —— 若后续页面补全,需回填

🔗 关联

⚠️ 语料说明:英文引句均为官方原句,未作改动(the AI 托管产品名为去品牌化形态)。

输入至少 2 个字符,检索全部课程与单词表