Multi-agent system, single agent or conventional automation?
Five criteria decide between conventional automation, a single AI agent and a multi-agent system: whether the steps are the same every time, whether the input is structured or free text, how many kinds of information the decision combines, how costly a mistake is, and how far the decision must be explained. If the steps are fixed, conventional automation is enough. If steps must be chosen case by case, consider a single agent. If different specialisms need to check one another, consider a multi-agent design.
This piece is a decision aid. We define the three approaches elsewhere: a single agent in what is an AI agent, the parts of a multi-agent design in what is a multi-agent AI system, and where rule-based flows begin in automation for small businesses. Here the subject is how to decide which is right for your own work.
The three approaches in one sentence each
- Conventional automation: a flow that runs on pre-written rules, in the same order every time. “When an order is confirmed, raise the invoice and email the customer.”
- Single agent: one piece of AI software that decides, case by case, which tool to use and in what order to reach a goal.
- Multi-agent system: several agents, each focused on a narrow job, producing a decision together under an orchestrator.
Think of them as rungs on a ladder. Each rung adds flexibility, and in return adds cost, testing effort and oversight.
Five criteria
1. Are the steps the same every time?
This is the first and most decisive question. If the steps can be listed in advance and run in the same order for every case, conventional automation is the cheapest, fastest and most dependable option. For work like this, the only thing an agent adds is unpredictability.
If the steps vary — one request needs the order record first, another the courier’s system first — then steps must be chosen. That is an agent’s job.
2. Is the input structured or free text?
If the input comes from a form, a spreadsheet or defined fields sent by another system, rules can process it directly. If it is free text — an email, a complaint, a contract — its meaning has to be extracted first.
There is an in-between option that is often overlooked: a single model step inside conventional automation. The flow stays rule-based; only the question “which category is this email?” or “what date does this document give?” goes to a language model. This is not an agent: the model chooses no steps, it simply fills in a field. Most business needs are met by this in-between design.
3. How many kinds of information does the decision combine?
If the decision rests on one kind of information — only the customer’s message, only one document — a single agent is enough. If it rests on a mix of numerical data, text, historical records and outside sources, giving each kind of information to a separate specialist builds a more consistent and testable structure.
Agents are divided by kind of information or specialism. A system doing ten similar jobs may well need just one agent.
4. How costly is a mistake?
If a wrong output is easily put right — a wrong label, a draft to edit — a simple structure is enough and the mistake is caught at human review. If a mistake is expensive — money moving, a binding reply to a customer, a record that cannot be reversed — splitting the decision into parts that check one another, and adding a stopping point, become important.
As the cost of error rises, the first remedy is usually a human approval step. One way to classify steps for approval is set out in hallucination, data privacy and human approval.
5. Does the decision need to be explained?
If you must be able to answer “why was this decided?” — to a customer, an auditor or your own team — the structure has to record it. Conventional automation is explainable by nature: the rule is there to read. In a single agent, every step must be logged with its reasoning. In a multi-agent design, each view feeding the decision is recorded separately, so you can see which part contributed to which outcome.
The decision table
| Criterion | Conventional automation | Single agent | Multi-agent system |
|---|---|---|---|
| Steps | Fixed, countable | Vary case by case | Vary, and need more than one specialism |
| Input | Structured (free text possible with a model step) | Free text, documents, mixed | Several kinds of source |
| Kinds of information | One or few | One main kind | A combination of different kinds |
| Cost of error | Low, or bounded by rules | Medium; managed with human approval | High; parts check one another |
| Need to explain | The rule is the explanation | Step log needed | Log of views and decision needed |
| Build and maintenance effort | Lowest | Medium | Highest |
The rule for using it: start from the column most of the five rows point to — but if you can stay in a simpler column, stay there. If only one row pulls to the right, that row’s problem can usually be solved without enlarging the whole structure. If only the cost of error is high, for example, the answer is an approval step.
Example tasks and the suggested structure
| Task | Suggested structure | Why |
|---|---|---|
| Raising an invoice and emailing it when an order is confirmed | Conventional automation | Fixed steps, structured input |
| Routing support emails to teams | Automation plus one model step | Fixed flow; only classification needs free text |
| An assistant answering from company documents | Single agent or a RAG assistant | One main kind of information, limited step choice |
| Checking order and courier systems for a customer request and drafting a reply | Single agent | Steps vary with the request |
| Pre-assessing applications against documents, past records and outside sources | Multi-agent system worth considering | Different kinds of information, high cost of error, need for reasons |
How an assistant that works from company documents is built is covered in what is RAG.
When do you move up a rung?
Starting with the simplest structure and moving up only when measurement calls for it is safer than building something complex from day one. The signs:
- From automation to a model step: the rules grow new exceptions every month and still cannot cope with free-text input.
- From a model step to a single agent: the steps change with each request, and a fixed flow no longer covers every case.
- From a single agent to a multi-agent design: the single agent’s brief has grown so wide that results are inconsistent in testing; it performs differently on different kinds of information, and you cannot tell which information an error came from.
Before every move, the same test set is run. If the new structure does not beat the old one on the same criteria, the move is not made. How to build the test set is covered in how to evaluate an AI system.
Common mistakes
- Handing a fixed process to an agent. For work that follows the same steps every time, an agent adds cost, delay and unpredictability.
- One agent per task. Agents should be divided by kind of information or specialism. A separate agent for every small task means needless connections and maintenance.
- Trying to solve the cost of error with architecture. The first remedy for a risky step is an approval step. A new agent does not replace approval.
- Moving up without measuring. “Multi-agent will work better” is an assumption. The decision rests on results measured on the same test set.
Frequently asked questions
Is a multi-agent system always better than a single agent?
No. More parts mean more model calls, more testing and more maintenance. It earns its place where different kinds of information meet, mistakes are expensive and decisions must be explained. Elsewhere, a single agent or conventional automation is the better choice.
Can conventional automation and AI be used together?
Yes, and most business projects are built that way. The flow stays rule-based, and the one step that needs to understand free text goes to a language model. That in-between design draws on what AI is good at without the cost and unpredictability of an agent.
Which structure costs least?
Conventional automation costs least to build and to run. A single agent adds usage cost because it makes several model calls per request. A multi-agent design needs the most effort both to build and to operate. We break down the cost items in what determines the cost of an AI project.
If we cannot decide, where should we start?
With the simplest structure, in a pilot built around one task. The pilot’s measurements show whether the next rung is needed. How to set one up is covered in how to start an AI project in your company.
If you would like to weigh up which structure suits your work, see our AI systems architecture or automation pages, or write to us.