One assistant against three: when a chain of assistants works better
One good chat for every task works up to a certain volume; beyond it the context interferes and the result drifts. Below is an analysis of a live chain where each assistant does its own part, and five standard pairings that cover most tasks.
One good chat for every task works up to a certain volume; beyond it the context interferes and the result drifts. Below is an analysis of a live chain where each assistant does its own part, and five standard pairings that cover most tasks.
One chat for everything is a bottleneck
A person masters working with assistants and does the natural thing: finds a good one and does everything in it.
It works — up to a certain volume. Then the symptoms appear.
The context interferes. In one conversation you discussed a client, and in the next message you ask for a text — and the model drags the previous thing in.
The result drifts. The same type of task comes out differently depending on how you happened to phrase it this time.
Quality is mediocre everywhere. There is no depth in any part of the task, because the assistant is holding style, structure, logic and grammar in mind at once.
The problem is not the model. You are using one universal tool where a set of specialised ones is needed.
What a chain of assistants is
The task is broken into parts, each with its own performer, and each result feeds the next.
Not "do everything for me" but a sequence where each works within its own frame.
A live example from practice. A series of covers in a single style is needed.
First — the art director. The task: how a series is built so the frames read as one line. It produced a split into constants and variables: a modular grid and an anchor zone, one typographic voice, a running colour anchor at a 60-30-10 proportion, consistent framing — those do not change. The subject figure, the secondary colour, the text hook — those do.
Second — the prompt generator. It received that brief and turned it into a generation prompt: a two-column modular grid at 9:16, exact colour codes — a graphite background, muted olive, a mustard-gold accent — the bottom right 15% left empty for the duration overlay.
What happened here. The art director does not know the syntax of generative models and should not have to. The generator does not know your brand system and should not have to. Each did its own part, and not one requirement was lost in the handover.
Ask one assistant to "make a series of covers" and you get pictures with no system.
Why handing over works better
Three reasons, all practical.
Each works within its own frame. An assistant does not have to think about meaning and form at the same time.
Requirements are fixed in text. When you write the prompt yourself, half the decisions stay in your head — "well, it should be contrasty". In a handover between roles everything is written down.
The intermediate result can be checked. If you do not like the brief, you correct that rather than redoing the final result.
Standard chains
Five pairings that cover most tasks.
Meaning → form. A specialist assistant supplies the content, a copywriter turns it into text, an editor cleans it.
Analysis → proposal. One assistant analyses a conversation with a client and extracts the pains, a second assembles a deck brief from them.
Data → conclusion → material. An analyst takes the figures apart, a marketing one turns the conclusions into content.
Draft → critique → correction. One writes, a second reviews it against criteria, the first redoes it against the notes.
Role → checked by another role. The legal one looks at the wording, the sales one at how it sounds to a client.
The last pairing is the most underrated: a check by another role finds what the author cannot see by definition.
Modes that replace a second role
Separately, because almost nobody knows about them.
Most assistants have built-in refinement modes. After an answer they offer:
Assess the work — a review of their own result against criteria with scores and reasoning.
Run a harsher assessment.
Give feedback as a focus group — the model goes through the result with the recipient's eyes: a buyer, a client, a reader.
Give an expert assessment.
Refine against the notes — corrections against criteria rather than a general "make it better".
Propose a new approach or change the format, style, length.
The practical point. You do not rewrite the request ten times. You get the first answer, run a harsh assessment, then a focus group — and end up with a document that has been through three rounds of checking. Without your involvement beyond choosing the mode.
For many tasks that replaces a second role in the chain.
How to assemble your own chain
An order in which this does not turn into extra work.
Step 1. Break the task into different kinds of work. Not into stages of time but into genuinely different types: invent, produce, check.
Step 2. Find a role for each. The catalogue has 14 categories: marketing, sales, creative, media, legal, HR, analytics, e-commerce, engineers, entrepreneurs, personal development, PR, client retention, freelance.
Step 3. Hand over the result whole. Not a summary but the text of the previous step. Summarising loses exactly what you started the chain for.
Step 4. Check at the joins. The most common place for loss is the transition between roles.
Step 5. Do not lengthen it without need. Two or three roles cover most tasks. A chain of six assistants usually means the task was badly divided.
When a chain is not needed
Honestly, because complicating for the sake of it is a common mistake.
A simple task. One question, one answer. There is no need to build a conveyor to write a letter.
One-off work. Setting up a chain pays back on repetition.
When there is no intermediate result you can judge. If you do not understand whether the first produced a good brief, handing it to the second will improve nothing.
What this gives in practice
Predictability. One type of task comes out the same way, because the order is set.
Transferability. A chain can be handed to an employee or a client along with a description.
Accumulation. A worked-out pairing stays with you and gets reused. The second client costs less than the first.
And scale. If the task recurs, a chain of assistants becomes an automation scenario — in n8n and Make templates (Creator plan and above) the result of each step is passed automatically.
What the platform offers for a chain of assistants
Over 200 AI assistants across 14 categories. Each is a role with its own frame and its own clarifying questions.
Refinement modes on most: assessment, harsh assessment, focus group, expert review, criteria-based refinement.
n8n and Make templates — if the pairing recurs: a scenario where the handover is automatic. Standard templates come with the Creator plan; the descriptions are open to everyone.
Creating your own assistants — the Creator+ plan, when the ready roles are not enough.
Chat access from the Basic plan. Registration is free and opens three days of full Basic.
Where to start today
One action. Take a task you usually do in one chat and split it into two different kinds of work.
For example: invent and produce. Or analyse and write.
Run it through two assistants, handing the first result to the second whole, with no summarising.
The difference is usually visible at once — and it explains why one universal assistant was producing mediocre results.
How this works in multi-agent systems
A chain of two or three assistants is the manual version. Once there are many roles, a hierarchy appears.
In multi-agent systems three levels of agent are usually distinguished:
Director — the top level, distributing tasks and consolidating the result.
Supervisor — the middle, running its own area and managing the performers.
Basic — the performer of a specific operation.
An example of such a hierarchy is the AI Company — Director template: a set of 8 n8n workflows with role agents (CFO, CTO, COO, Assistant, Client Finder and others). Standard templates come with the Creator plan.
Why know this if you work in chats. The same logic applies with no automation at all: one assistant sets the frame, a second performs within it, a third checks. The only difference is that in a system the handover is automatic.
The templates — Creator, Make, N8N and Full.
The chain as a ready tool
n8n and Make automation templates — when a chain recurs, it is convenient to assemble it as an automatic scenario: the steps are connected and one result becomes the next one's input. Standard templates come with the Creator plan, advanced ones with Full. For content tasks there is Cerebrum: describe the task in plain words, the right models are picked for you, paid from the balance on actual use, the price visible before launch.
A ready example of a seven-link chain — "Freelancer positioning", $9 or the Basic plan. Seven assistants where each result becomes the next one's input: Personality unpacker → Skill map → Digital brand architect → Career editor → Strategic consultant → Creative packager → Landing page architect.
That is the best illustration of this article: it shows why one universal assistant could not do it — each link has its own frame and its own questions.
Training. The “AI Agents: the electrification of routine” express course, $9 one-off or the Creator, Make, N8N and Full plans: four platforms for different tasks — Make for non-coders, n8n for full control, Relevance and Lindy for ready scenarios.
Skills for assembling a chain
All six are advanced skills: Creator and above, Full, or $9 one-off or 1000 XP.
- Designing multi-agent systems — precisely this article's subject: how to divide the work, who to assign what, how to pass context between steps and consolidate results
- Task coordination strategies — four ways to divide a task: by layer, by component, by cross-cutting responsibility, by file ownership. Plus an assignment template with acceptance criteria and boundaries
- Task complexity classification — it answers this article's main question quantitatively: it scores a task out of ten and says where one assistant is enough and where a team is needed
- Orchestration with parallel performers — three phases: a strong model plans, simple ones perform in parallel, the strong one checks against criteria. A maximum of two attempts at refinement
- Delegating to sub-agents — and a practical reason for chains: the main context does not get cluttered. The assistant reads the files itself and returns a short summary rather than a wall of text
- Context control in chains — a continuation of the same thought: what not to use so that assistants' transcripts do not eat the window
What determines whether a chain is needed at all. The third item. Not "does it look difficult" but a scored assessment of the task: below a certain level a chain adds more overhead than it gives in quality.
What to take on the platform for this task
- Art director and GPT Prompt Master — the assistants from the covers example: one sets the system of the series, the other turns it into a prompt; from the Basic plan
- AI copywriter — an assistant for the “meaning → form” link; from the Basic plan
- AI Company — Director — an n8n template: role agents in one system; Creator plan and above
- Swarms — multi-agent orchestration framework — Open Source: the description is open to all, the repository link from Creator
- AI Agents: the electrification of routine — an express course; $9 one-off or the Creator, Make, N8N and Full plans
What to read next
Six brief fields — what to give the first in the chain as input.
How to build your own AI assistant — when the ready roles are not enough.
A virtual board of directors — a chain of six roles on one decision.
Skills for Claude Code — the same logic inside one tool.