Manager + Writers — How Smart-Copy.ai's Multi-Agent Architecture Generates 150-Page Documents

Why ChatGPT Can't Write a Coherent 50-Page Document

Anyone who's tried to generate a longer text in ChatGPT knows the problem: after a few thousand words, the model starts losing its way. It repeats earlier fragments, forgets what it wrote at the beginning, loses the main narrative thread. You ask for a continuation — and the next part sounds like it was written by a completely different person. This isn't ChatGPT's fault or any other AI model's — it's a fundamental limitation of large language model (LLM) architecture, which has a finite "context window" and simply cannot maintain coherence in very long texts.

Smart-Copy.ai solves this problem in a way that seems simple at first glance but actually required months of R&D: instead of one AI model writing the entire text, we use a team of specialized agents who collaborate like a magazine editorial team. The Manager plans the structure and oversees everything, while Writers execute individual sections according to the Manager's guidelines. In this article, we'll show exactly how this architecture works — step by step, with technical details and examples.

The Context Window Problem — Why One Model Isn't Enough

Before explaining the solution, it's worth understanding the problem. Language models like Claude or GPT have a limitation called the "context window" — the maximum amount of text they can "see" at once. For Claude Sonnet 4.5, this is about 200,000 tokens (roughly 150,000 words), which sounds impressive. The problem is that this window must accommodate the system prompt, sources, instructions, AND the generated text. In practice, much less space remains for the actual output.

But even if the context window were infinite, there's a second problem: generation quality drops with very long texts. The model "forgets" earlier decisions, repeats itself, loses stylistic consistency. This isn't a bug — it's a feature of the transformer architecture that all modern LLMs are based on. A single model simply isn't designed to write 50-page documents in one go. A different approach is needed.

Limits in Practice

When generating a 10-page article (approximately 20,000 characters) in a single API call, the model receives: system prompt (~5,000 tokens), research sources (~30,000-50,000 tokens), structure and SEO instructions (~2,000 tokens). Only about 10,000-15,000 tokens remain for the actual output text — meaning a maximum of those 20,000 characters. But what if you need 50, 100, or 150 pages? A single model physically cannot generate that.

Multi-Agent Architecture — A Team Instead of an Individual

The solution is multi-agent architecture, inspired by how real editorial teams and content marketing departments work. In Smart-Copy.ai, we've implemented a two-level system: the Manager is responsible for planning and coordination, while a team of up to 7 Writers handles the actual writing. Each agent is a separate AI model call with a dedicated prompt and task. Agents communicate through structured data — the Manager passes the plan to Writers, Writers return completed sections.

This architecture solves both problems simultaneously. First, each Writer generates only a text fragment (typically 3,000-8,000 characters), so they fit comfortably in the context window with full access to sources and instructions. Second, the Manager maintains overall coherence — knows the full structure, monitors what's already been written, and gives Writers precise guidelines on what to write next. Result: a 50-page text is just as coherent as a 5-page one.

The Manager — The Brain of the Operation

The Manager is the first agent that launches after the research phase completes. Its task is to create a complete article plan based on: the topic and guidelines from the user, selected research sources, required text length, SEO parameters (keyword, internal links). The Manager doesn't write any content — its output is a structured plan in JSON format.

The Manager's plan contains: a list of all sections (H2 and H3 headings), estimated length of each section in characters, key points to cover in each section, placement of formatting elements (where tables, lists, examples should go), assignment of keywords and SEO links to specific sections. The Manager also calculates how many Writers will be needed — for text up to 20,000 characters one is enough, for longer texts the system assigns additional Writers (up to a maximum of 7).

Example Manager Plan

For an article on "Content Marketing for E-commerce — Complete Guide" with a required length of 25,000 characters, the Manager might generate the following plan:

Section	Heading	Length	Writer	Elements
1	H2: Why E-commerce Needs Content Marketing	~3,500 ch	Writer 1	Statistics, keyword in H2
2	H2: Types of Content for Online Stores	~4,000 ch	Writer 1	Bullet list, examples
3	H2: Store Blog — How to Write to Sell	~5,000 ch	Writer 2	H3 subsections, SEO link #1
4	H2: Product Descriptions That Convert	~4,500 ch	Writer 2	Comparison table, examples
5	H2: SEO for E-commerce — Basics	~4,000 ch	Writer 3	SEO link #2, checklist
6	H2: Measuring Content Marketing Results	~3,000 ch	Writer 3	KPI table, summary

Writers — Content Specialists

After the Manager approves the plan, the system launches Writers sequentially — one after another. Each Writer receives a dedicated prompt containing: the full article plan from the Manager (so they know the overall context), research sources, specific sections to write (e.g., "write sections 3 and 4"), the last 5,000 characters of text written by the previous Writer (for smooth transitions), a list of already-written headings (preventing repetition).

The Writer generates only their assigned sections — they don't try to write the entire article. This is crucial for quality: instead of 25,000 characters in one call (which would exceed comfortable limits and reduce quality), each Writer generates 6,000-10,000 characters with full access to sources and clear context. The quality of each section is therefore as high as in a short article.

Context Passing Mechanism

The most important element of the architecture is how Writers "know" what their predecessors wrote. Each subsequent Writer receives:

• Last 5,000 characters of the previous section — this allows for smooth stylistic and narrative transition. The Writer "sees" how the previous part ends and can naturally continue.
• List of all written H2 and H3 headings — this prevents topic repetition. If Writer 1 wrote about "types of content," Writer 2 won't duplicate it.
• Summary of key points — if specific numbers or definitions appeared in section 2, subsequent Writers can reference them consistently.
• Position information in the text — the Writer knows whether they're writing the middle of the article (can reference earlier and preview later sections) or the ending (should summarize).

Self-Repair System — When Things Go Wrong

In a multi-agent system, things can go wrong: the model might generate shorter text than expected, cut off mid-sentence (token limit), skip a required element (e.g., SEO link), or generate a duplicate heading. Smart-Copy.ai has built-in self-repair mechanisms that automatically detect and correct these problems.

Each Writer's output goes through a validator that checks: whether the text has the required length (±15% tolerance), whether the last sentence is grammatically complete, whether all required SEO elements were placed, whether there are no heading duplications. If validation detects a problem, the system automatically launches a correction — for example, an additional AI call with the instruction "finish the last sentence and add a summary" or "insert the missing SEO link naturally."

Continuation Mechanism for Truncated Text

The most common problem is text truncation due to token limits — the model simply stops generating at a random point. The system detects this automatically (the last character isn't a period, exclamation mark, or question mark ending a sentence) and launches a continuation. The Writer receives a prompt: "Continue exactly from where you stopped: [last 500 characters]. Complete the current thought and section." The system can perform up to 5 such continuations before considering a section complete.

Dynamic Scaling — From 1 to 7 Writers

The number of Writers isn't fixed — the system dynamically selects it based on required text length. The algorithm is simple: each Writer can comfortably generate 5,000-10,000 high-quality characters. For a 20,000-character text, 2-3 Writers are enough; for 50,000 characters, 5-6 are needed; and for the maximum 300,000 characters (about 150 A4 pages), the system launches a full team of 7 Writers in multiple rounds.

The Manager calculates the optimal number of Writers during planning and distributes sections evenly. This ensures no Writer is overloaded (which would reduce quality) and the entire text generates in optimal time. For a 10-page article (1 Writer), this is 2-4 minutes; for 50 pages (4-5 Writers) — 8-15 minutes; for 150 pages (7 Writers, multiple rounds) — 30-60 minutes.

Text Length	Number of Writers	Generation Time	Example Use Case
Up to 15,000 characters	1	2-4 min	Blog article, product description
15,000 - 40,000 characters	2-3	4-8 min	Comprehensive guide, report
40,000 - 80,000 characters	4-5	8-15 min	Ebook, whitepaper
80,000 - 150,000 characters	5-6	15-30 min	Extensive ebook, documentation
150,000 - 300,000 characters	7 (multiple rounds)	30-60 min	Book, compendium

Stylistic Consistency — How to Maintain It

A skeptic might ask: if different Writers (different AI calls) are writing the text, won't it be stylistically inconsistent? This is a valid concern — and exactly why the Manager plays such an important role. Besides planning structure, the Manager defines a "style guide" for the entire article: tone (formal, conversational, expert), difficulty level (for laypeople, specialists, experts), formatting (how often to use lists, tables, examples), narrative voice (first person, third person, impersonal).

Each Writer receives the same style guide in their prompt, plus a tone example from previous sections (last 5,000 characters). In practice, stylistic consistency is very high — often higher than in texts written by one person over several days, because the AI model doesn't have "bad days" and doesn't forget the established tone between sessions. In A/B tests, readers couldn't identify where one Writer's work ended and another's began.

Comparison with Competition — Why This Works Better

Most AI generators on the market use simpler approaches: they either generate text in one call (with length and quality limitations), or offer "continuation" — the user must manually ask for the next part of the text, and the model generates it without full context of previous sections. Smart-Copy.ai is the only generator we know of that implements full multi-agent architecture with a Manager and Writer team.

Aspect	ChatGPT / Claude	Jasper / Copy.ai	Smart-Copy.ai
Maximum length	~4,000 words / response	~5,000 words / document	~150 pages (300,000 characters)
Long text coherence	Drops with continuations	Medium	High (Manager + context)
Automatic planning	None	Basic templates	Full plan from Manager
Context passing	Manual (copy-paste)	Limited	Automatic 5000 ch + headings
Error self-repair	None	None	Validation + auto-continuation
User effort	High (prompts, stitching)	Medium	Minimal (form)

Technical Implementation Details

For those interested in technical details: Smart-Copy.ai is built on the Claude Sonnet 4.5 model from Anthropic. Each agent (Manager and Writers) is a separate API call with a dedicated system prompt. Communication between agents occurs through structured JSON data — the Manager returns a plan in JSON Schema format, which is validated before being passed to Writers. The entire pipeline is implemented in TypeScript on Fastify, with PostgreSQL as the database and queuing through the order system.

Sequential Writer execution (not parallel) is intentional — each subsequent Writer needs the previous one's output for context. This extends generation time compared to a parallel approach, but dramatically improves coherence. We tested a parallel variant and quality was noticeably lower — Writers "stepped on each other's toes," repeating similar content without knowing what others were writing.

Summary — An Orchestra Instead of a Soloist

The multi-agent architecture in Smart-Copy.ai is the answer to fundamental limitations of single AI models. Instead of forcing one model to write a 50-page document (which it can't do well), we orchestrate a team of specialized agents: the Manager plans, Writers write, validators check, the self-repair system corrects errors. The result is texts up to 150 A4 pages long, with stylistic and substantive coherence comparable to work from a professional editorial team.

This isn't a marketing slogan — it's a real technological advantage that no other generator on the market offers. ChatGPT will cut off after 4,000 words. Jasper requires manual section stitching. Smart-Copy.ai will generate a complete ebook without any user intervention — from one click to a finished document.

Want to see the multi-agent architecture in action? Create a free account on Smart-Copy.ai and order a longer text — for example, 30,000 or 50,000 characters. Watch how the system automatically selects the number of Writers and generates a coherent, professional document. Starting at 3.99 PLN per 1,000 characters, no subscription.