When one AI agent is enough and when you need a team of agents.
A single AI agent is like a very capable newcomer who rarely says ‘I don't know’. It answers quickly, convincingly and with the same confidence whether it is right or wrong. So the deciding question is not whether the agent makes mistakes, but what happens when it makes one.
An agent won't say: ‘I'm not sure’.
Language models are trained to produce answers that sound plausible. They only partly know the limits of their knowledge and admit uncertainty unreliably. So they often deliver a wrong answer as convincingly as a correct one. Newer models are better at this, but it won't disappear completely, because it follows from how models work.
For an email summary or a draft post, it doesn't matter. For a contract, a payment or a reply to a patient, it is a risk.
With people
Companies figured this out long ago. Larger payments are signed by two people. With a serious diagnosis, you get a second opinion. A pilot has a co-pilot and a checklist. Not because people can't do their jobs, but because nobody else would catch the mistake of a single approver.
With AI
You can build the same wisdom into the system. Instead of one agent that answers, you build a small team that checks each other.
A team of agents works on the principle of more eyes.
A good example is the Apollo 11 mission control. Before every critical decision, the flight director asked each specialist: go or no go? One ‘no go’ was enough to stop. A system of agents can be built the same way:
- The first agent prepares the answer. Quickly, as a first draft.
- The second agent verifies it. It checks facts and looks for made-up data.
- The third agent looks for mistakes. Its job is to break the proposal and ask what could go wrong.
When they agree, you can trust the result more. When they disagree, it is a signal: the proposal goes to a person, not out.
“A law firm lets an agent prepare a first draft of a reply to a client. A second agent compares the cited provisions with the text of the law and the contract. A third looks for wording that could bind the client to something unwanted. If any of them disagrees, the draft lands with a lawyer with a note on what doesn't fit.”
More agents cost more. It only pays off in some places.
Each additional agent means more model calls and higher costs. So it shouldn't be the solution to everything. A simple test helps: what happens when AI gets it wrong?
| CONSEQUENCE OF A MISTAKE | SOLUTION |
|---|---|
| Minor inconvenience (bad recommendation, weaker text) | One agent, simple and cheap |
| Money, health, legal risk, regulation | A team of agents with verification and a person when they disagree |
A main agent, yes. Keys to everything, no.
The second question is whether to have one ‘main’ agent for the whole company. It makes sense. People don't want to look for which chat is for invoices and which for sales. They want one place that understands the whole company and knows where to send a task.
The risk arises when such an agent gets all the rights. It can read, change and delete everywhere. If someone misuses or manipulates it through one weak point, they get to everything. The damage then spreads from one place to the whole company.
With people
A receptionist is one point of contact for the whole company. She knows whom to pass what to. But she doesn't have the keys to the safe, payroll or the server room. Only those who need them do.
With AI
The main agent is the receptionist. It understands the request and passes it to a specialized agent. Only one narrow agent that does nothing else has access to a specific system, for example accounting.
A secure design therefore rests on three principles:
- As few permissions as possible. Each agent gets only as much as it needs for its task. The riskier the system, the narrower the agent.
- Separate tools. Only one path leads to a sensitive system. If something goes wrong, the damage stays in one place.
- Log everything. Every decision and every step is recorded so you can trace what happened and why.
And above it all: a person. Either directly in the process for important decisions, or at least with an overview of what the agents do.
When not to use a team of agents.
If an agent recommends articles to read or prepares internal summaries, a team of agents is an unnecessary expense. If the company doesn't yet have a single agent that works reliably, start with one. A team makes sense where a mistake hurts and where you already know exactly what to check.
One agent or a team: how to decide
An agent is wrong just as convincingly as it is right. So the question isn't whether to trust it, but what happens when it gets something wrong.
Where a mistake costs nothing, one agent is enough. It is cheap and fast. Where a mistake costs money, health or reputation, build a team: one agent proposes, a second verifies, a third looks for mistakes, and when they disagree, a person decides.
And if you want one main agent for the whole company, let it work like a receptionist. It understands the request and passes it on, but only narrow agents with minimal rights hold the keys to sensitive systems. You get the convenience of one place without the risk that one mistake hits the whole company.
Ask yourself
Which decision would you entrust to an agent only if someone checked it afterwards?
