This analysis organizes the evidence behind designing a Multi-Agent Content Production System, then explains the practical implications, trade-offs and current limits.
Read the evidence below as a decision trail: what changed, why it matters, which trade-offs shaped the result, and where the conclusion still depends on context.
Use three full-time AI Agents to establish an enterprise content production line: article writers are responsible for knowledge output, network information experts are responsible for information transcription, and website architects are responsible for technical maintenance. The entire process from material to online is automated, and the production time of a single article is reduced from 3-6 hours to 30-60 minutes.
The real bottleneck in enterprise content production is not “no one writes”
There are two types of content needs within the enterprise that have been handled inefficiently for a long time.
The first isInternal knowledge output. Department members have a lot of experience and decision-making logic, but the cost of converting from "ideas in their heads" to "communicable documents" is extremely high. An in-depth article of 2000-3000 words takes 3-5 hours to write manually. The result is that valuable knowledge remains at the individual level and cannot be reused by the team.
The second isexternal intelligence gathering. There are dozens of important releases in the AI industry every week scattered across more than a dozen English sources. It takes 3-6 hours to track, read, understand, and convert them into actionable intelligence one by one. The approach of most teams is to "look at it when you have time and take it as what you see", which lacks systematicity.
The common essence of these two problems:The cost of converting raw information into structured output is too high.It’s not that no one has ideas, it’s not that there are no sources worth tracking, but the middle conversion link—structuring, arrangement, translation, formatting—eats up more than 80% of the time. And these links happen to be what AI is best at doing.
Why use three Agents instead of one?
The solution is not to "find an AI to write articles", but to design a multi-agent division of labor architecture. Three full-time Agents perform their own duties:
article-writer: Receive scattered materials (notes, ideas, experiences) from department members and transform them into structured and argumentative articles. Responsible for extracting core arguments, planning article structure, writing content, fact checking, and automatic publishing.
Internet information expert (web-intelligence): Automatically scan 14 English sources, filter relevant articles, deconstruct and reorganize them into Traditional Chinese for in-depth analysis. Responsible for source monitoring, topic selection suggestions, content transcription, fact checking, and automatic listing.
Website Architect (gwarket-site): Maintain website technical architecture (Next.js static website generation (SSG) + WordPress headless content management system (Headless CMS) + Cloudflare Pages). Responsible for performance optimization, architecture decisions, bug fixes, and deployment management.
The three Agents share the same publishing pipeline: automatic publishing through WordPress REST API, triggering Cloudflare Pages Deploy Hook to automatically rebuild and go online.
The judgment logic of splitting
Why not have one Agent do everything?Different responsibilities require different context definitions. Website architects need to understand Next.js and Cloudflare technology stacks; article writers need to understand writing style and structural specifications; information experts need to understand source lists and transliteration logic. Stuffing them all together will overload the context and degrade quality.
Why not tear down more?For example, proofreading and publishing are independent? Proofreading requires seeing the original material for comparison. If it is taken apart, the context will be lost. Putting on the shelf is a mechanical action and does not require an independent role. It is more reasonable to make a shared command. Excessive splitting will increase manual handling costs.
This splitting principle is the same as that of enterprise organization design:Split according to functional boundaries, not according to process steps.Each Agent's CLAUDE.md (skill package) is independent, and modifying one does not affect the others - it is the same concept as the modular design of software engineering.
How to draw the line of division of labor between man and machine
There is only one design principle:AI does tasks that are highly repetitive and have clear rules. People make strategic judgments and quality checks.
Taking article writers as an example, the division of labor is as follows:
- Material reception: AI extracts core arguments and actively asks questions to supplement; humans provide original materials and opinions
- structural planning: AI plans the article structure; humans confirm the direction
- content writing: AI produces complete articles and SEO metadata; people review content quality
- fact check: AI automatically compares materials and detects the information it fills in; humans judge marked items
- On the shelves: AI is automatically released through API; people only need to say "put it on the shelves"
The key is the last check. "To send or not to send" is a strategic decision, not a technical judgment, so it is left to people. But the tedious work of checking itself — comparing data item by item, scanning sources of proper nouns — is left to AI.
AI will make mistakes, the key is to use mechanisms to correct them rather than relying on humans to make them up.
Problems encountered in actual operation and solutions:
AI factual bias.When the article writer produced the article, he wrote the deployment platform Cloudflare Pages as Vercel - because Next.js was mentioned in the material, AI automatically guessed it based on common combinations. After discovery, we not only correct the errors, but also add an "AI fill-in detection" mechanism to the review process: AI is required to scan all proper nouns before delivery. Those that are not explicitly mentioned in the material are marked as warnings and proactively ask for confirmation.Use mechanisms to plug loopholes instead of expecting the AI to make no mistakes next time.
Performance deteriorates.After adding Featured Image, Lighthouse dropped from 88 to 74. Because verification is run at every step, it was discovered in time that the 2560px original image was loaded as a thumbnail. Instead, use the medium size thumbnail returned by the WordPress API to solve the problem.
Deploy Hook is triggered repeatedly.WordPress triggered rebuild twice when updating the post. Use WordPress transient mechanism to add anti-duplication logic correction.
WordPress API is blocked.The Build server called the API and returned a 415 error. After tracing, it was found that the Wordfence security plug-in was interfering. The settings were adjusted and retry logic was added.
Common patterns among these problems:Instead of relying on people to guard against errors, we should establish mechanisms to allow errors to be automatically detected and systematically repaired.
How to cut MVP: only do the shortest path from "material to online"
What has been done in the first version: definition of skill packages for three Agents, complete pipeline for article writers and information experts (material → article → review → listing), shared shelving automation, fact checking mechanism (including AI filling detection), GEO optimization format.
What the first version doesn’t do:
- Pictures are automatically generated — The quality of AI output images is unstable, and manual settings are more controllable.
- social posts — Insufficient ROI, focus on website content quality first
- Automatically scheduled scans — Manually triggering the verification process is sufficient
- Multi-language output — Focus on Traditional Chinese first, English version will be added in the future
- Collaborative review —Currently a single person operates, but multi-person processes increase complexity.
Every exclusion is because itDoes not affect whether the core process can run smoothly. The cutting logic of MVP is to focus on the shortest path - first prove that the line "material → article → online" can operate automatically, and then talk about expansion.
The efficiency of a single article is increased by 3-8 times, and the time it takes to put it on the shelf is reduced from 15 minutes to 1 minute.
- Internal knowledge output: 3-5 hours/article → 30-60 minutes/article, efficiency increased by 3-5 times
- External information transcription: 3-6 hours/article → 25-45 minutes/article, efficiency increased by 5-8 times
- From shelf to online: 15-30 minutes of manual work → automatically completed in 1 minute, 15-30 times faster
- SEO Zero Loss: The URL structure remains unchanged, and there is no new 404 in Google Search Console.
After time is saved, manpower is reallocated to things that AI cannot do: strategic judgment, topic selection decision-making, and quality control.
What this means
This project allowed me to verify one thing:The introduction of AI into enterprise content processes is essentially an organizational design issue, not a technology procurement issue.
The approach of many companies is to buy an AI tool, let employees "try it out," and then shelve it when they find that the effect is not as good as expected. The problem is not that the tools are not good, but that the structure of human-machine division of labor has not been designed well - which links are handed over to AI, which AI is handed over, at what point humans intervene, and how to correct errors when errors occur. If these decisions are not made well, no matter how good a tool is, it is just a toy.
The value of the Multi-Agent architecture does not lie in "how many AIs are used", but inIt forces you to break down the process. When you want to define the responsibility boundaries and skill packages of each Agent, you must answer: What exactly is this link doing? What is the input? What is the output? What are the quality standards? Many companies have never clarified these issues during manual operations.
The verification mechanism of AI is more important than the output capacity of AI.Output quality can be improved iteratively, but without a systematic review mechanism, errors will be amplified—the faster AI produces, the faster errors will spread. "AI fill-in detection" in fact-checking is the product of this logic: instead of expecting AI to make no mistakes, it is better to design a mechanism to intercept errors before going online.
Practical questions and boundaries
Does an enterprise need an engineering team to introduce the Multi-Agent architecture?
Not necessarily. The three Agents of this project were all built using Claude Code. The core is the CLAUDE.md skill package definition of each Agent - essentially writing a specification document, not a program. What is needed is someone who can break down the process clearly, not necessarily an engineer. However, if technical links such as website architecture or API connection are involved, corresponding technical judgment is still required.
How to judge whether a certain link should be handed over to AI or left to humans?
Look at two dimensions: repeatability and judgment complexity. The highly repetitive and clearly defined links (structuring, formatting, comparison and verification) are handed over to AI. The links that require strategic judgment and value trade-offs (topic selection, tone setting, and decision-making announcements) are left to people. Gray areas are handled by letting AI do it first and humans reviewing it — for example, fact-checking is performed by AI, but flagged warning items are judged by humans.
How exactly does the AI filling detection mechanism work?
Before the article is delivered, AI will scan all proper nouns (company name, product name, platform name, tool name) in the article and compare them one by one with the original materials provided by the user. Nouns not explicitly mentioned in the material will be marked as "AI filling" and the user will be actively asked for confirmation. The design principle of this mechanism is: it is better to leave blanks for others to fill in than to have the AI guess wrong on its own.
What to take away
The article's value is in the evidence and trade-offs behind designing a Multi-Agent Content Production System, not in treating the conclusion as universal.