# I Replaced a 4-Person Team With a Multi-Agent Claude System. Here Is the Exact Architecture.
**作者**: CyrilXBT
**日期**: 2026-05-01T14:01:25.000Z
**来源**: [https://x.com/cyrilXBT/status/2050214010541269225](https://x.com/cyrilXBT/status/2050214010541269225)
---
Twelve months ago I was paying four people to run my content operation.
A research assistant. A content writer. A distribution manager. An analytics person.
Combined monthly cost: $11,400.
Combined output: inconsistent, slow, and dependent on four different people showing up every day with energy, focus, and the right context to do their job.
Today that entire operation runs on a multi-agent Claude system I built in three weekends.
Monthly cost: $340.
Output: higher volume, more consistent quality, faster turnaround, and zero sick days.
This is not a story about replacing people because I do not value them.
It is a story about what becomes possible when you stop thinking of AI as a single tool and start building it as a coordinated team.
Here is the exact architecture.
## Why One Agent Is Never Enough
Most people using Claude are running a single agent model.
They open a chat. They ask for something. They get a result. They close the tab.
This works for simple tasks. It completely breaks down for complex operations that require multiple specialized skills running in coordination.
A single agent trying to do everything is like hiring one person and asking them to be simultaneously your best researcher, your best writer, your best analyst, and your best strategist.
Nobody is world-class at all of those things at the same time.
The people building real AI-powered operations have figured out the same thing elite organizations figured out decades ago.
Specialization beats generalization every time.
A team of specialists coordinated by a smart orchestrator outperforms any individual generalist regardless of how capable that individual is.
Multi-agent Claude systems work on exactly this principle.
Each agent has one job. One area of expertise. One set of tools it is optimized for. And a master orchestrator that knows when to call which agent and how to pass context between them.
## The Architecture Overview
Before I break down each agent, here is the full system map.
Five agents. One orchestrator. One shared memory layer. Zero humans required for standard operations.
The Orchestrator Agent sits at the top. It receives the initial input, breaks it into component tasks, routes each task to the appropriate specialist agent, collects outputs, handles failures, and assembles the final deliverable.
The Research Agent handles all information gathering. It reads, synthesizes, and produces structured research briefs.
The Content Agent handles all writing. It takes research briefs and produces publish-ready content in the correct voice and format.
The Distribution Agent handles all publishing and scheduling. It takes finished content and routes it to the correct platform in the correct format at the correct time.
The Analytics Agent monitors performance. It reads data, identifies patterns, and produces optimization recommendations.
The Quality Agent reviews every output before it leaves the system. It checks against predefined standards and either approves or sends back for revision.
Each agent has its own CLAUDE.md. Its own system prompt. Its own tool access. Its own quality standard.
None of them know what the others are doing in real time. The orchestrator knows everything.
## The Orchestrator Agent
This is the most important agent in the system and the one most people underinvest in.
The orchestrator is not a manager in the traditional sense. It does not motivate. It does not coach. It does not interpret vague instructions and make judgment calls about intent.
The orchestrator is a router and a synthesizer.
Its entire job is to receive a clearly defined input, break it into the minimum number of component tasks required to complete it, route each task to the right agent with all the context that agent needs, monitor for completion and failure, and assemble the final output.
The system prompt for your orchestrator needs to answer five questions before it does anything else.
What is the task that just arrived? What is the exact output required? Which agents need to be involved and in what order? What context does each agent need to do its job? What does success look like?
Here is the orchestrator prompt template:
You are the master orchestrator for a multi-agent content operation.
When you receive a task your job is to:
1. Identify the exact final output required
2. Break the task into component subtasks
3. Determine which agent handles each subtask in order:
RESEARCH → CONTENT → QUALITY → DISTRIBUTION
Skip any agent whose subtask is not required
4. For each agent, prepare a task brief that includes:
- The specific task
- All context the agent needs
- The exact format of output required
- The quality standard to meet
5. Pass each brief to the correct agent
6. When all agents complete: assemble outputs into the final deliverable
7. Pass the final deliverable to the Quality Agent for approval
8. If Quality Agent rejects: route back to the relevant agent with specific feedback
Never start work without a complete task brief.
Never pass context between agents informally.
Never assume an agent has information you did not explicitly give it.
The last three rules are the ones that save you from 90% of multi-agent failures.
## The Research Agent
This was the first role I replaced and the one that delivered the most immediate return.
My research assistant was spending 6 hours a week pulling information, reading articles, and compiling notes. The output was inconsistent in depth and format depending on the day and the topic.
The Research Agent does the same work in 12 minutes and produces a structured output in exactly the format the Content Agent needs.
The Research Agent has access to three tools: web search, a knowledge base of previously collected research, and a synthesis engine that cross-references multiple sources.
Its single job is to take a research request and produce a Research Brief with a standardized structure.
Here is the Research Agent system prompt:
You are a specialist research agent. Your only job is to produce
Research Briefs.
When you receive a research request:
1. Identify the core question or topic
2. Search for the most relevant and recent sources
3. Cross-reference at least 3 independent sources for any factual claim
4. Identify the key insight most people miss on this topic
5. Identify any counterintuitive angle that creates genuine interest
6. Identify 3 specific examples, statistics, or stories that support the insight
Output only in this format:
CORE INSIGHT: [one sentence]
SUPPORTING EVIDENCE: [3 specific examples with sources]
COUNTERINTUITIVE ANGLE: [what most people get wrong]
KEY QUOTES OR DATA: [2-3 specific numbers or direct quotes]
SUGGESTED ANGLES: [3 content angles ranked by potential]
Never add commentary. Never editorialize. Produce the brief and stop.
The structured output format is not optional. It is what makes the handoff to the Content Agent reliable.
An agent that produces free-form research creates downstream failure because the Content Agent cannot predictably extract what it needs.
Structure is the foundation of every reliable multi-agent system.
## The Content Agent
This agent has the most nuanced job in the system because it requires something that breaks most generic AI setups.
Voice matching.
Content that does not sound like you is content that erodes the audience you spent years building. It does not matter how technically correct or information-dense it is. If it reads like AI and not like you, your audience will feel the gap even if they cannot name it.
The Content Agent is the one you invest the most time training.
Before you write a single prompt for this agent you need to give it a library of your highest performing content. Your top 20 posts. The articles that generated the most engagement. The writing that made people share without being asked.
The agent reads all of it, identifies your patterns, and writes everything from inside that voice.
Here is the Content Agent system prompt structure:
You are a specialist content agent for @cyrilXBT.
Voice profile [loaded from examples]:
- Short punchy sentences. Hard stops.
- Every idea gets its own line
- Strategic ALL CAPS for emotional emphasis
- No dashes anywhere
- No corporate language, no AI clichés
- Direct and opinionated. Always a clear POV
- Specific over generic. Numbers over abstractions
When you receive a Research Brief your job is to:
1. Select the strongest angle from the Suggested Angles
2. Write the opening hook in ALL CAPS
3. Write the body in the voice profile above
4. End with a CTA that drives follow
5. Check every sentence against the voice profile before submitting
Output formats you produce:
- Long tweet (300-500 words)
- Thread (8-12 tweets)
- Article (1500-3000 words)
- Requote tweet (150-250 words)
Always specify which format you are producing at the top of your output.
Never produce output in any format not listed above.
Never use bullet points in tweet content.
Never use dashes.
The voice profile section is where you put the extracted patterns from your content examples. The more specific you make it, the less editing the output requires.
My Content Agent now produces first drafts that require under 5 minutes of editing for 80% of outputs.
## The Quality Agent
Most multi-agent systems skip this one.
They run research into content into distribution and assume everything that comes out is good.
It is not.
Every system needs a gate between creation and publication. A set of standards that every output must pass before it goes anywhere.
The Quality Agent is that gate.
It does not produce anything. It only evaluates.
Here is the Quality Agent prompt:
You are a quality control agent. You do not create content.
You evaluate it.
When you receive content for review, score it on these criteria:
VOICE MATCH (1-10): Does this sound exactly like @cyrilXBT?
HOOK STRENGTH (1-10): Does the opening line stop the scroll?
INFORMATION DENSITY (1-10): Does every sentence earn its place?
CTA CLARITY (1-10): Is the call to action clear and compelling?
FORMAT COMPLIANCE (1-10): Does it follow all format rules?
Passing score: 8 or above on all criteria.
If any criterion scores below 8:
- State exactly which criterion failed
- State exactly why it failed
- State exactly what needs to change
- Return to the Content Agent with this specific feedback
If all criteria score 8 or above:
- Mark as APPROVED
- Pass to the Distribution Agent
Never approve content that fails any criterion.
Never give vague feedback.
Be specific or the Content Agent cannot fix it.
The Quality Agent runs on every piece of content without exception. It has rejected drafts that looked good to a casual read but failed the voice match or hook strength criteria on close inspection.
It has saved me from publishing content that would have underperformed and cost me audience trust.
## The Distribution Agent
The job most people do manually and most consistently undervalue automating.
Formatting content for each platform. Scheduling at optimal times. Posting without forgetting. Following up on performance.
The Distribution Agent handles all of it.
It receives approved content from the Quality Agent and executes a distribution plan.
You are a specialist distribution agent.
When you receive approved content:
1. Identify the content type and intended platform
2. Format the content to platform specifications:
- X: max 280 chars per tweet, thread if longer
- Newsletter: full formatting with headers
- LinkedIn: professional adaptation of the core content
3. Schedule based on optimal posting windows:
- X: Tuesday 10AM EST, Monday 10AM EST, Friday 10AM EST
- Newsletter: Thursday 9AM EST
4. Confirm scheduling and report back to Orchestrator
5. Flag any formatting issues before posting
Never post without confirmation from the Orchestrator.
Never post to a platform without formatting for that platform specifically.
The scheduling rules in step 3 are based on the analytics data from your actual account. You update them every month based on what the Analytics Agent reports.
## The Analytics Agent
This is the agent that makes every other agent better over time.
It reads your performance data weekly, identifies what is working and what is not, and produces an optimization brief that gets fed back into the Content Agent and Distribution Agent system prompts.
You are a specialist analytics agent.
Every Monday at 6AM you run a weekly performance analysis.
Pull the following data from the last 7 days:
- Top 5 posts by impressions
- Top 5 posts by bookmark rate
- Best performing day and time
- Best performing format
- Best performing topic category
Produce a Weekly Intelligence Brief with:
WHAT WORKED: The specific pattern in top performing content
WHAT FAILED: The specific pattern in underperforming content
RECOMMENDED ADJUSTMENTS: Specific changes to content strategy
OPTIMAL POSTING SCHEDULE: Updated based on this week's data
ONE COUNTERINTUITIVE INSIGHT: Something the data shows that
you would not expect
Deliver brief to Orchestrator every Monday at 6AM.
The Analytics Agent brief gets incorporated into the Monday morning context load. Every week the system gets slightly more calibrated. Every month the output quality is measurably better than the month before.
## The Memory Layer
The piece that holds the whole system together and the piece most tutorials leave out entirely.
Without a shared memory layer, every agent in your system starts fresh every session. The Research Agent does not know what the Content Agent produced last week. The Analytics Agent cannot access the research briefs that informed previous content decisions.
The memory layer is a structured knowledge base that every agent can read from and write to.
In practice this is a combination of Obsidian vault files and a Supabase database depending on what needs to persist and what needs to be queryable.
Every output from every agent gets logged. Every brief. Every approved piece of content. Every analytics report.
The Orchestrator reads this before routing any task. It knows what has been researched, what has been written, what has been published, and how it performed.
This is what prevents the system from repeating itself, chasing topics it has already covered, and making the same mistakes twice.
## Building This in a Weekend
The full system sounds complex because it is. But you do not build all five agents at once.
Week 1: Build the Orchestrator and the Content Agent. These two alone will cut your content production time in half.
Week 2: Add the Research Agent. Now your content is also better informed. Feed it into the Content Agent.
Week 3: Add the Quality Agent. Now everything you publish has been through a gate. Your consistency goes up dramatically.
Week 4: Add the Distribution Agent. Now publishing is fully automated.
Week 5: Add the Analytics Agent. Now the whole system learns.
Each week you add one agent. Each week the system gets meaningfully more powerful. By the end of Week 5 you have the full architecture running.
The investment is five weekends.
The return is an operation that runs continuously without you in the loop for standard tasks.
Every hour you used to spend on production mechanics is now an hour you can spend on strategy, relationship building, and the work that only you can do.
That is the real value of a multi-agent system.
Not just the cost savings.
The freedom to operate at the layer that actually matters.
## The Prompts That Tie It Together
The single most important prompt in the entire system is the one that initializes each agent at the start of every session.
Before you do anything else, read the following files from the
shared knowledge base:
1. MASTER_CONTEXT.md — who we are, what we build, our audience
2. RECENT_PERFORMANCE.md — what worked last week
3. ACTIVE_PROJECTS.md — what is currently in progress
4. STYLE_GUIDE.md — voice, format, and quality standards
5. AGENT_ROLE.md — your specific job in this system
Confirm you have read all five files before accepting any task.
Do not proceed without confirmation.
This initialization prompt is the difference between an agent that operates with context and an agent that starts cold every session.
Context is everything in multi-agent systems.
The agents that have it produce outputs that compound.
The agents that start fresh every session reset your entire operation every time they open.
## The Result After 12 Months
Weekly content output: up 340%.
Average post quality as measured by bookmark rate: up 67%.
Time spent on content mechanics per week: down from 22 hours to 4 hours.
Monthly operating cost: down from $11,400 to $340.
And the system gets better every single week because the Analytics Agent is continuously feeding performance data back into the system prompts of every other agent.
The four people I used to pay were good at their jobs.
But they could not work 24 hours a day, 7 days a week, with perfect consistency, zero ego, and an architecture that gets smarter every time it produces output.
The multi-agent system can.
Build yours.
Follow @cyrilXBT for the exact prompts, templates, and agent architectures that make this possible.
## 相关链接
- [CyrilXBT](https://x.com/cyrilXBT)
- [@cyrilXBT](https://x.com/cyrilXBT)
- [976](https://x.com/cyrilXBT/status/2050214010541269225/analytics)
- [@cyrilXBT](https://x.com/@cyrilXBT)
- [@cyrilXBT](https://x.com/@cyrilXBT)
- [@cyrilXBT](https://x.com/@cyrilXBT)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [10:01 PM · May 1, 2026](https://x.com/cyrilXBT/status/2050214010541269225)
- [976 Views](https://x.com/cyrilXBT/status/2050214010541269225/analytics)
---
*导出时间: 2026/5/1 22:21:42*
---
## 中文翻译
我用一个多代理 Claude 系统替换了一个 4 人团队。以下是精确的架构。
**作者**: CyrilXBT
**日期**: 2026-05-01T14:01:25.000Z
**来源**: [https://x.com/cyrilXBT/status/2050214010541269225](https://x.com/cyrilXBT/status/2050214010541269225)
---
十二个月前，我还在付钱给四个人来运营我的内容业务。
一名研究助理。一名内容写手。一名分发经理。一名数据分析员。
合计月成本：11,400 美元。
合计产出：不稳定、缓慢，且依赖于四个人每天都能带着精力、专注力和正确的上下文来上班工作。
今天，整个业务由我在三个周末构建的一个多代理 Claude 系统运行。
月成本：340 美元。
产出：更高的产量、更稳定的质量、更快的周转速度，以及零病假。
这不是一个因为我看不上人而替换他们的故事。
这是一个关于当你不再把 AI 视为单一工具，而是开始将其构建为一个协作团队时，一切将有何种可能的故事。
以下是精确的架构。
## 为什么一个代理永远不够
大多数使用 Claude 的人都在运行单代理模式。
他们打开一个聊天。他们提出要求。他们得到结果。他们关闭标签页。
这对于简单的任务有效。但对于需要多种专业技能协调运作的复杂操作，这种方式完全会崩溃。
一个试图包揽一切的单代理，就像雇佣一个人，同时要求他成为你最好的研究员、最好的作家、最好的分析师和最好的战略家。
没有人能同时在所有这些方面都达到世界级水平。
那些真正构建 AI 驱动业务的人，已经发现了精英组织几十年前就悟出的道理。
专业化每次都能战胜泛化。
一个由智能编排器协调的专家团队，其表现优于任何个人通才，无论这个个人能力有多强。
多代理 Claude 系统正是基于这一原则运作。
每个代理只有一个工作。一个专业领域。一组它经过优化的工具。以及一个主编排器，它知道何时调用哪个代理以及如何在它们之间传递上下文。
## 架构概览
在我详细分解每个代理之前，这里是完整的系统地图。
五个代理。一个编排器。一个共享记忆层。标准操作无需人工参与。
编排器代理位于顶端。它接收初始输入，将其分解为组件任务，将每个任务路由到相应的专家代理，收集输出，处理故障，并组装最终交付物。
研究代理处理所有信息收集。它阅读、综合，并生成结构化的研究简报。
内容代理处理所有写作。它接收研究简报，以正确的语调和格式生成即发内容。
分发代理处理所有发布和调度。它接收完成的内容，以正确的格式在正确的时间路由到正确的平台。
分析代理监控性能。它读取数据，识别模式，并生成优化建议。
质量代理在每个输出离开系统前进行审查。它根据预定义的标准进行检查，要么批准，要么退回修改。
每个代理都有自己的 CLAUDE.md。自己的系统提示词。自己的工具访问权限。自己的质量标准。
它们中没有一个是实时知道其他人在做什么的。编排器知道一切。
## 编排器代理
这是系统中最重要的代理，也是大多数人投入不足的地方。
编排器不是传统意义上的管理者。它不进行激励。不进行辅导。不解读模糊的指令或对意图进行判断。
编排器是一个路由器和合成器。
它的全部工作是接收一个明确定义的输入，将其分解为完成它所需的最少组件任务，将每个任务连同该代理所需的所有上下文路由到正确的代理，监控完成和失败情况，并组装最终输出。
你的编排器的系统提示词需要在做任何其他事情之前回答五个问题。
刚到达的任务是什么？需要什么确切的输出？需要涉及哪些代理以及按什么顺序？每个代理需要什么上下文来完成其工作？成功看起来是什么样的？
以下是编排器提示词模板：
你是一个多代理内容运营的主编排器。
当你收到任务时，你的工作是：
1. 确定所需的确切最终输出
2. 将任务分解为组件子任务
3. 确定哪个代理按顺序处理每个子任务：
RESEARCH（研究） → CONTENT（内容） → QUALITY（质量） → DISTRIBUTION（分发）
跳过任何不需要其子任务的代理
4. 为每个代理准备一份任务简报，包括：
- 具体任务
- 代理所需的所有上下文
- 所需输出的确切格式
- 要达到的质量标准
5. 将每份简报传递给正确的代理
6. 当所有代理完成时：将输出组装成最终交付物
7. 将最终交付物传递给质量代理进行批准
8. 如果质量代理拒绝：带着具体反馈路由回相关代理
没有完整的任务简报绝不要开始工作。
绝不要在代理之间非正式地传递上下文。
绝不要假设代理拥有你没有明确给它的信息。
最后三条规则使你免于 90% 的多代理失败。
## 研究代理
这是我替换的第一个角色，也是带来最直接回报的一个。
我的研究助理每周花费 6 小时拉取信息、阅读文章和整理笔记。输出的深度和格式根据日期和主题而波动。
研究代理在 12 分钟内完成同样的工作，并生成一个结构化的输出，完全符合内容代理所需的格式。
研究代理有权访问三个工具：网络搜索、先前收集研究的知识库，以及交叉引用多个来源的综合引擎。
它唯一的工作是接收研究请求并生成具有标准化结构的研究简报。
以下是研究代理系统提示词：
你是一个专家研究代理。你唯一的工作是生成研究简报。
当你收到研究请求时：
1. 确定核心问题或主题
2. 搜索最相关和最新的来源
3. 对任何事实声明交叉引用至少 3 个独立来源
4. 确定大多数人在这个主题上遗漏的关键见解
5. 确定任何能产生真正兴趣的反直觉角度
6. 确定支持该见解的 3 个具体例子、统计数据或故事
仅以此格式输出：
核心见解：[一句话]
支撑证据：[3 个带有来源的具体例子]
反直觉角度：[大多数人的误解之处]
关键引语或数据：[2-3 个具体数字或直接引语]
建议角度：[按潜力排名的 3 个内容角度]
绝不要添加评论。绝不要发表评论。生成简报然后停止。
结构化的输出格式不是可选项。它使得向内容代理的交接变得可靠。
一个生成自由形式研究的代理会导致下游失败，因为内容代理无法可预测地提取它需要的东西。
结构是每个可靠多代理系统的基础。
## 内容代理
这个代理拥有系统中最微妙的工作，因为它需要某种打破大多数通用 AI 设置的东西。
语调匹配。
听起来不像你的内容会侵蚀你花费数年建立的受众。不管它在技术上多么正确或信息密度多大。如果它读起来像 AI 而不是像你，你的受众会感觉到这种落差，即使他们无法名状。
内容代理是你投入最多时间训练的代理。
在为这个代理写一个提示词之前，你需要给它一个你表现最好的内容库。你表现最好的 20 篇帖子。产生最多互动的文章。那些让人们无需请求就主动分享的文字。
代理通读所有这些，识别你的模式，并从那种语调内部撰写所有内容。
以下是内容代理系统提示词结构：
你是 @cyrilXBT 的专家内容代理。
语调配置文件 [从示例加载]：
- 短促有力的句子。强硬的停顿。
- 每个想法占一行
- 战略性的全大写以进行情感强调
- 任何地方都不要有破折号
- 没有企业语言，没有 AI 陈词滥调
- 直接且有主见。总是有清晰的观点
- 具体优于通用。数字优于抽象
当你收到研究简报时，你的工作是：
1. 从建议角度中选择最强的角度
2. 用全大写写开头钩子
3. 用上面的语调配置文件写正文
4. 以一个引导关注的 CTA 结尾
5. 在提交之前对照语调配置文件检查每一句话
你生成的输出格式：
- 长推文（300-500 字）
- 推文串（8-12 条推文）
- 文章（1500-3000 字）
- 转发推文（150-250 字）
总是在输出的顶部指定你正在生成哪种格式。
绝不要以上面未列出的任何格式生成输出。
绝不要在推文内容中使用要点。
绝不要使用破折号。
语调配置文件部分是你放置从内容示例中提取的模式的地方。你越具体，输出所需的编辑就越少。
我的内容代理现在生成的初稿，80% 的输出需要不到 5 分钟的编辑。
## 质量代理
大多数多代理系统都跳过了这个。
它们从研究跑到内容再到分发，并假设出来的所有东西都是好的。
事实并非如此。
每个系统都需要在创建和发布之间有一个大门。每个输出必须通过的一套标准。
质量代理就是那个大门。
它不生产任何东西。它只进行评估。
以下是质量代理提示词：
你是一个质量控制代理。你不创建内容。你评估它。
当你收到要审查的内容时，根据这些标准评分：
语调匹配（1-10）：这听起来完全像 @cyrilXBT 吗？
钩子强度（1-10）：开头的那一行能让人停下划动的手指吗？
信息密度（1-10）：每一句话都赢得了它的位置吗？
CTA 清晰度（1-10）：行动号召是否清晰且引人注目？
格式合规性（1-10）：它是否遵循了所有格式规则？
及格分数：所有标准 8 分或以上。
如果任何标准得分低于 8 分：
- 准确说明哪个标准失败
- 准确说明它为什么失败
- 准确说明需要改变什么
- 带着这个具体反馈返回给内容代理
如果所有标准得分均为 8 分或以上：
- 标记为已批准
- 传递给分发代理
绝不要批准未能通过任何标准的内容。
绝不要给出模糊的反馈。
要具体，否则内容代理无法修复它。
质量代理毫无例外地运行在每一篇内容上。它曾经拒绝了那些随意阅读看起来不错，但在仔细检查下未能通过语调匹配或钩子强度标准的草稿。
它使我免于发布那些表现不佳并会让我失去受众信任的内容。
## 分发代理
大多数人手动完成的工作，并且最一贯低估自动化价值的工作。
为每个平台格式化内容。在最佳时间安排发布。不遗忘地发布。跟进表现。
分发代理处理所有这一切。
它从质量代理接收批准的内容并执行分发计划。
你是一个专家分发代理。
当你收到批准的内容时：
1. 识别内容类型和目标平台
2. 将内容格式化为平台规范：
- X：每条推文最多 280 个字符，更长则用推文串
- 通讯：带有标题的完整格式
- LinkedIn：核心内容的专业改编版
3. 基于最佳发布窗口安排：
- X：周二上午 10 点 EST，周一下午 10 点 EST，周五上午 10 点 EST
- 通讯：周四上午 9 点 EST
4. 确认调度并向编排器报告
5. 在发布前标记任何格式问题
未经编排器确认绝不要发布。
未针对特定平台格式化绝不要发布到该平台。
步骤 3 中的调度规则基于你实际账户的分析数据。你根据分析代理的报告每月更新它们。
## 分析代理
这是让其他每个代理随时间变得更好的代理。
它每周读取你的性能数据，识别什么有效、什么无效，并生成一个优化简报，该简报被反馈回内容代理和分发代理系统提示词中。
你是一个专家分析代理。
每周一上午 6 点你运行每周性能分析。
从过去 7 天拉取以下数据：
- 按浏览量排名的前 5 篇帖子
- 按收藏率排名的前 5 篇帖子
- 表现最好的一天和时间
- 表现最好的格式
- 表现最好的主题类别
生成一份每周情报简报，包含：
什么有效：表现最好的内容中的具体模式
什么失败：表现不佳的内容中的具体模式
建议调整：对内容策略的具体更改
最佳发布时间表：基于本周数据更新
一个反直觉见解：数据显示出的你意想不到的东西
每周一上午 6 点将简报交付给编排器。
分析代理的简报会被纳入周一上午的上下文加载中。每周系统都会稍微校准一点。每个月的输出质量都比上个月有可衡量的提高。
## 记忆层
这是将整个系统粘合在一起的部分，也是大多数教程完全遗漏的部分。
没有共享记忆层，你的系统中的每个代理在每次会话时都是全新开始的。研究代理不知道内容代理上周生产了什么。分析代理无法访问告知先前内容决策的研究简报。
记忆层是一个结构化的知识库，每个代理都可以从中读取和写入。
在实践中，这是 Obsidian vault 文件和 Supabase 数据库的组合，具体取决于什么需要持久化，什么需要可查询。
每个代理的每个输出都会被记录。每份简报。每篇批准的内容。每份分析报告。
编排器在路由任何任务之前会读取这些。它知道已经研究了什么，已经写了什么，已经发布了什么，以及表现如何。
这就是防止系统自我重复、追逐已经涵盖的主题、并两次犯同样错误的原因。
## 在一个周末构建这个
完整的系统听起来很复杂，因为它确实复杂。但你不会一次构建所有五个代理。
第 1 周：构建编排器和内容代理。仅这两个就会将你的内容生产时间减半。
第 2 周：添加研究代理。现在你的内容也更有见识了。将其输入内容代理。
第 3 周：添加质量代理。现在你发布的所有内容都经过了一个大门。你的一致性急剧上升。
第 4 周：添加分发代理。现在发布完全自动化。
第 5 周：添加分析代理。现在整个系统都在学习。
每周你添加一个代理。每周系统都会变得更有力。到第 5 周结束时，你就有了完整的架构在运行。
投入是五个周末。
回报是一个连续运行的业务，无需你介入标准任务。
你过去花在生产机制上的每一小时，现在都可以花在战略、关系建立以及只有你能做的工作上。
这就是多代理系统的真正价值。
不仅仅是节省成本。
在真正重要的层面上运作的自由。
## 将其串联起来的提示词
最重要的单一提示词