Vibe 编程：多智能体工程
停止只和一个 AI 助手工作，开始领导一个智能体团队。为已经上线生产产品的创始人编写，他们现在想以真实工程团队的速度进行构建。你将学会创建专化的 AI 智能体：一个产品经理、一个架构师、一个开发者、一个评审者、一个安全专家、一个 DevOps 工程师，给每一个赋予明确的角色，并像现代软件团队一样协调他们。你指挥团队，智能体去执行工作。到最后，你用完整的 AI 工程团队而不是单个助手来运行一个完整的软件项目。
共 12 节，按顺序
- 120 分钟
Thinking Like an Engineering Leader
在前两门课里，你学会了做出一个产品，并把它交付给真实用户，靠的全是跟一个 AI 对话。这门课要请你换一把椅子坐。你不再是自己带着一个助手把每一步都做完，而是要组建并带领一支 AI 智能体（agent）团队，也就是你早就在用的同一个...
An engineering workflow is the repeatable path work takes: plan, build, review, shipA team splits ownership so no single owner holds everythingA leader invests in a plan before any code existsA non-technical leader makes technical calls by weighing the tradeoffs the AI lays out - 220 分钟
Creating Specialized AI Agents
上一课你换了位置，从动手写代码的人变成了负责规划和协调的工程负责人。这一课要认识的，是你将要带的这支团队。它仍然是一节只读的课，没有代码，也不用打开工作区。它要做的事情只有一件：把组成一支人工智能工程团队的角色名单摆在你面前，让你在开始指挥之前先弄清楚谁负责什么。
An AI agent is the same AI given one focused role, its own instructions, and its own slice of the workA team of specialists beats one generalist doing everythingThe seven roles of an AI engineering team and what each one ownsA role instruction is the short system-style prompt that defines an agent's job - 320 分钟
Giving Every Agent a Role
上一课你设计了一支团队：产品经理、架构师、开发者、评审者，还有其他几个角色，每一个都要交给同一个 AI 去扮演。不过，列出一串角色还算不上一支团队。只有当每个角色都定义得足够精确，让智能体 (agent)...
Responsibilities: naming what an agent owns and, just as important, what it does not ownContext: the goal, constraints, and prior agents' decisions that must travel with a taskRole instruction: the standing description that makes an agent behave the same way every timeHandoffs: passing work between agents without losing informationTurning a vague role into a precise one - 430 分钟
Managing AI Collaboration
前面三课里，你设计了一支智能体（agent）队伍，给每一个分了角色，也学会了用干净的交接在它们之间传活儿。现在，你要让这支队伍同时上手产品里的同一块东西。这一课的目标是：让几个智能体一起把一个功能做出来，彼此不踩脚，而你坐在中间当协调者。这件事你要在工作区里通过给智能体写提示词来完成，不用写任何代码...
Shared context: the one common understanding all agents work from (the goal, the plan, what is done)Planning: agreeing the approach before anyone starts buildingCommunication: relaying one agent's output to the next as clean inputReviews: one agent checking another's work before it counts as doneIteration: looping build, review, refine instead of one-shotting a featureDriving one small feature through planner, developer, and reviewer as the coordinator - 530 分钟
AI Code Reviews
上一课里，你让一个小功能依次经过了规划者、开发者和评审者，也看到了为什么评审者必须是另一个独立的智能体。这一课要把评审者变成一门真正的规矩。代码评审指的是一个智能体去读另一个智能体的成果，在它进入产品之前先把不对的地方揪出来，这是一个团队能养成的性价比最高的质量习惯。你会在工作区里跑一个评审者，读它...
The review process: how a change goes from proposed to reviewed to merged, and why nothing merges unreviewedWhy the reviewer must be a different agent from the author, so it reads with fresh eyesQuality checks: whether the work meets the requirement, is clear, and has nothing missing or brokenRefactoring: improving the structure of working code without changing what it doesSecurity reviews: flagging unsafe patterns such as trusting user input without checking itActing on findings as the leader, deciding what to fix without reading the code line by line - 620 分钟
AI Project Management
到目前为止，你已经学会带一支智能体团队：给每个智能体一个专注的角色，在它们之间交接工作时给足上下文，并在信任任何成果之前单独跑一遍评审。这是把活儿干好的方式。这一课讲的是你作为负责人同时还担着的另一份差事：把项目本身管起来。不只是指挥每个智能体，还要决定团队接下来做什么、按什么顺序做，以及你怎么把这...
The leader runs the project, not just the individual agentsA sprint is a short fixed period with a committed set of work; sprint planning is choosing that setTask management breaks a plan into tracked tasks with one owner and a checkable doneDocumentation is a written record of decisions and how things work so the team stays alignedPriorities decide what matters most now and are revisited as things changeUsing a PM agent to draft a sprint plan, a task list, and priorities - 730 分钟
AI DevOps
你已经带着一队智能体把产品做出来、评审过，也把项目管起来了。在前面那门课里，你还照着步骤亲手把它上线过一次。这一课要换的，是"上线"这件事归谁管。以后不再由你把每个新版本一个个推出去，而是把交付和运行整个交给 DevOps...
Handing the delivery and running of the product to a DevOps agent so shipping is automatic and repeatableCI/CD as an automatic pipeline that tests and ships every changeDeployments as releasing new versions safely and repeatably, including the ability to roll backInfrastructure described as config the team can review, not servers set up by handMonitoring where the DevOps agent wires up health checks and alerts so problems surface fastThe founder directs the DevOps agent and reviews what it produces - 830 分钟
AI Security Teams
现在你的产品已经摆在真实用户面前，还有一个 DevOps 智能体负责把它发出去。第二门课里你学过保障产品安全的基本功，比如给密码做哈希，那时候这些都是你在开发过程中顺手做的事。这一课要把安全单独拎出来，让它在团队里占一个正式席位。你会组建一个安全智能体，一个...
A dedicated security agent whose only job is to find ways the product could be attackedA security review as a structured pass that reports weaknesses ranked by severityActing on findings by severity, worst first, instead of fixing everything at onceA vulnerability as a specific weakness an attacker could useA vulnerability scan as systematically hunting for known weaknesses, including in dependenciesA threat analysis as thinking like an attacker: who would target this, what they want, and where they would push+1 - 930 分钟
AI Operations
你的产品已经上线，有真人在用了。这件事改变了你要干的活儿。做产品是把新东西造出来；做运维是在别人依赖你的时候，把你已经交付出去的东西维持在健康状态。这一课里，你要往团队里再加一个职责明确的角色，一个运维智能体，并学会带着它熬过每个创业者都害怕的那一刻：生产环境出问题了。你不用写任何代码。你负责指挥、...
The operations agent as a focused role that keeps the live product healthyIncident response as the calm, ordered way to detect, diagnose, fix, and communicate when something breaks in productionMonitoring as watching signals so incidents are caught early, building on the monitoring you set up in course 2Production support as handling the steady stream of real-world issues and user reportsMaintenance as the ongoing upkeep that prevents incidents: updates, cleanup, and keeping things healthyThe blameless postmortem as a write-up after an incident of what happened and how to prevent it - 1020 分钟
Building Your Own AI Workflow
这门课程一路走来，现代工程团队需要的每一个角色你都见过了：决定做什么的产品经理、决定各部分怎么拼在一起的架构师、把它们做出来的开发者、检查成果的评审者、找弱点的安全专家，还有负责上线和运维的 DevOps...
An AI software factory is a repeatable path an idea travels from head to shippedWorkflow design is fixing the sequence of agent roles every feature passes through, in the same order each timeAutomation is handing repeated steps to agents or scripts so the leader stops doing them by handDocumentation is writing the workflow itself down so it is repeatable and shareable, not held in memoryProductivity is how fast good work flows through the line, improved by removing friction rather than by cutting reviewsAn end-to-end example: idea, PM, architect, build, review, security, deploy - 1120 分钟
Scaling AI Teams
你已经学会了带一支智能体团队：给每个智能体一个明确的角色，交接工作时把上下文一并带过去，另外单独安排评审，再用冲刺和优先级把项目管起来。只要项目小到一个智能体能整个装进脑子里，这一整套都跑得漂亮。这一课要谈的是：当项目长过了那个尺寸，什么东西变了；以及为什么拖慢你的不再是某一个智能体，而是这支团队怎...
A large project outgrows a single agent's context and must be split into parts with clear boundariesA boundary defines what a part owns and the one well-defined way other parts interact with itTeam coordination keeps many agents on the same goal without conflicting or duplicating workA shared source of truth holds docs and decisions outside any single conversation so context does not vanishLong-term development means onboarding new agents, guarding against drift, and keeping the plan currentCoordination and shared knowledge, not raw building speed, are the real limits at scale - 1245 分钟
Becoming an AI Engineering Leader
这是进阶课程的最后一课，也是整条三门课路径的最后一课。你从“AI 不是魔法”开始学起，学会了做出一个产品，再把它发布到生产环境，然后不再依赖单个助手，而是带起一支各有专长的智能体团队。现在，是把这一切合到一处的时候了：你要带着一整支 AI...
Assembling the full AI engineering team and giving each agent its roleThe seven roles: product manager, architect, developer, reviewer, security, DevOps, and operationsRunning a project through the whole workflow: plan, build, review, secure, ship, operateHandoffs that carry decisions and context, not just the next instructionLeading the team rather than doing the work, and holding the quality standardCarrying the whole approach into the founder's own real projects
用你的语言学习 Vibe 编程：多智能体工程
每一节课、测验和导师回复都以你的语言显示。AI 导师会用你提问的语言来回答。