10 AI agents for multiple e-commerce sites

One agent for each area of work

The system includes ten agents. The most visible roles cover several stages of the e-commerce operation:

The other roles are not necessary to understand the general design. The system stays readable because every agent has a domain it needs to know well, and a set of actions it does not need to own.

This avoids turning one agent into a giant prompt that has to understand products, campaigns, SEO, stock and personal habits at the same time. One agent may be good at writing and poor at catalog management. Another may need access to Meta Ads without having any reason to read personal email.

That logic is consistent with the work I do on AI agents and workflow automation for companies: the model comes after the process. First, define who reads what, who prepares what and who is allowed to act. Depending on the level of autonomy, an agent such as Hermes can then orchestrate part of that work.

The brand expert becomes a central component

In a multi-agent architecture, brand knowledge is often more important than a particularly sophisticated prompt.

The brand expert knows the products, user manuals, important specifications and the tone to use. It can act as a reference point for the other agents when they need to produce copy: a product page, a blog post, a landing page, an advertising campaign or a social post.

That does not mean the brand agent should do everything. Its role is to provide the context that keeps another agent from inventing an argument or describing a product inconsistently.

The distinction matters:

• the e-commerce agent handles store information and surfaces;
• the SEO agent works on content structure and visibility;
• the visual agent produces the required assets;
• the brand expert protects consistency across the system.

Without that separation, every agent ends up with its own version of the brand. The texts may be correct when read separately, yet contradict one another when compared.

One conversation can involve several agents

Not every exchange stays isolated. In some cases, several agents take part in the same conversation.

A simple example:

1. The entrepreneur asks for a landing page for a product.
2. The brand expert drafts a first version from the available information.
3. The SEO agent reviews the copy and suggests improvements.
4. The e-commerce agent applies the result to the store or publishes the content in the right place.

The first agent does not need to know everything or execute everything. It can ask the relevant specialist to step in, then continue with the result in the same conversation.

That is the difference between a collection of independent assistants and an agent system. In the first case, the user has to copy and paste work from one tool to another. In the second, agents can pass a structured request between them, provided that their roles and limits are clearly defined.

Delegation does not remove controls. A page can be well written and poorly optimised. A text can match the brand voice and contain outdated product information. A post can be ready without needing to go live immediately. The chain therefore needs to show when the system is preparing, changing or publishing something.

Discord is the work interface

The agents are mainly used through Discord, with one-to-one conversations organised into threads. This choice solves a very practical problem: keeping one channel from becoming an unreadable history of every store, campaign and personal task.

A thread keeps the subject visible and prevents its history from being mixed with other tasks. The user can open one conversation with the e-commerce agent for a product page, another with the personal agent for a calendar task, then bring several agents together when a request crosses multiple domains.

Discord is not the control system, though. It is the interface the user sees. Source data, permissions, action logs and recovery mechanisms still need to be defined in the connected tools and in the project architecture.

That distinction matters for companies that simply want to add a bot to a channel. A discussion channel does not solve catalog duplicates, stock errors, bad advertising audiences or the publication of unchecked content. It makes the work more natural. It does not replace business rules. The same caution applies to the technical base: a self-hosted stack with Hetzner, Coolify and n8n can provide control, but it also creates an operating responsibility.

MCP provides tool access, not a general authorisation

The Ads agent is connected to Meta Ads through MCP. This type of connection lets the agent work in an environment that goes beyond generating text.

But MCP does not decide what the business allows the agent to do. Several levels still need to be separated:

• reading campaigns and their settings;
• preparing a change;
• changing an existing campaign;
• launching a campaign or changing a budget;
• stopping a campaign and explaining why.

These actions do not carry the same risk. A technical connection can work while the business permission is still poorly defined. That is exactly the kind of boundary I describe in the article about human approval for AI agents in production.

The same reasoning applies to the e-commerce agent. Editing a product page draft does not commit the business in the same way as changing a price, a stock figure or a page visible to customers.

A subscription does not replace a compliant architecture

In the configuration described here, the whole setup runs around a ChatGPT Pro subscription budgeted at €200, at the Pro 20X level. That amount belongs to this project's setup. It is not a Kirako service price or a general estimate for the cost of a multi-agent system.

The number alone does not make the architecture valid for every use. The official OpenAI documentation on ChatGPT Pro tiers, checked on September 18, 2026, reminds users that the plan is subject to limits around automated or programmatic extraction, credential sharing, making the account available to other people and using ChatGPT to power a third-party service. For an automated or multi-user deployment, verify the provider-approved access path and plan before reproducing this setup. A ChatGPT subscription is not a generic license for exposing ten agents to users or an external service.

Most of the agents use GPT-5.6 Sol, with Terra and Luna for certain needs. Not every role needs the same model capability, speed or way of handling a request.

The number of agents is therefore not enough to calculate cost or complexity. You also need to look at connected tools, image and video generation, hosting, platform access, monitoring and maintenance. Model choice deserves a separate review, as I explain in the comparison of open and proprietary AI models.

The right split is not necessarily the one that creates the most agents. It is the one that improves quality, safety or clarity. A dedicated SEO agent makes sense if it genuinely reviews the content. It only adds noise if it repeats the brand agent's instructions.

What to define before reproducing the setup

Before creating ten agents, I would start with five questions:

1. What is the reliable source? Product data, stock, manuals and brand rules should not be copied into multiple places without a clear owner.
2. Which actions are allowed? Separate reading, preparation, editing, publishing and advertising spend.
3. When is human approval required? A simple rule is better than a vague "be careful" added to a prompt.
4. How can the team see what happened? It should be possible to identify the request, the agent involved, the tool used and the result.
5. What happens when something goes wrong? A publication, stock change or campaign needs a correction procedure and, where possible, a way back.

These questions will last longer than the exact list of models used today. Models will change. Responsibilities, permissions and critical data will remain at the centre of the system.

What I take away from this project

The project is designed to make work transferable between specialists while keeping a clear boundary around actions that commit the store, the brand or the advertising budget.

For an entrepreneur running several sites, the first project will not necessarily be a ten-agent system. It can start with one loop: take reliable product information, produce a consistent page, have it reviewed and decide exactly who is allowed to publish it.

Once that loop works, the other agents become easier to add. You know what data they need to read, which actions they can prepare and where a human must take over. That is when a collection of prompts becomes a real AI agent system connected to the company's work. To define the sources, permissions and first workflow, an AI audit is often more useful than installing ten agents immediately.