What is a multi-agent AI system? What we learnt building CAI

A multi-agent AI system is a structure in which several “agents”, each focused on a single job, work together — usually under an orchestrator that weighs their outputs and turns them into one decision. It differs from a single model in that expertise, reasoning and limits are split into parts you can test, log and stop. You need one when a decision draws on different kinds of information, when conditions change, when mistakes are expensive or when the decision has to be explained.

Most AI projects start with one large model. You ask a question; the model answers. That is enough to summarise a document or draft an email. But once a decision is involved — money, risk, sequence, priority — expecting one model to do everything builds something brittle. Multi-agent systems are an answer to that brittleness. Below we explain what they are, when they are needed and what we learnt building our own autonomous trading system, CAI.

A multi-agent system, in brief

An agent is a piece of software with a defined task, defined data to look at and a defined output. Behind it might sit a language model, a statistical model or a set of simple rules. If the term is new to you, our article on what an AI agent is covers it from the ground up.

Above the agents there is usually an orchestrator. It gathers their outputs, decides how far to trust each one when they disagree, and makes sure the system arrives at a single result.

An analogy: asking one model for everything is like leaving every decision in a company to a single person. A multi-agent system is a table of specialists — and a chair who runs the meeting.

How does it differ from a single model?

There are three important differences:

• Expertise. Because each agent works in a narrow field, it produces more consistent and more auditable results. Testing an agent that answers “is the market trending or compressing right now?” is far easier than testing a model that is meant to know everything.
• Reasoning. Which agent’s view led to which decision is recorded. Later, the question “why was this decided?” has an answer.
• Limits. You can halt a risky step at a single point, make certain decisions subject to human approval, and budget cost per agent.

There is a price, too: more parts, more connections and a more careful architecture. Not every project needs a multi-agent structure.

When do you need a multi-agent system?

In our experience, a single model falls short in these situations:

1. When the decision combines different kinds of information — for instance, numerical data, text and historical records all at once.
2. When conditions change. If one approach works in one environment and another works elsewhere, splitting them across separate specialists is sturdier.
3. When mistakes are expensive. If a wrong decision costs a lot, dividing it into parts that check one another — and building in a way to stop — is essential.
4. When the decision must be explained — to a client, an auditor or yourself.

If none of these apply, a single well-written model call is often enough. Making the architecture more complex than it needs to be is a mistake of its own.

CAI: a ten-agent trading system

CAI (Compass AI) is an autonomous crypto futures system developed in our studio, running in production with real capital. Classic trading bots rest on a single strategy and weaken when the market changes character — from trend to compression, from calm to sudden volatility. With CAI, instead of a single “hero strategy”, we sought consensus from ten specialist agents.

The architecture has three layers:

• Perception. Agents that read the market’s current regime, on-chain signals and macro conditions. This layer makes no decisions — it only describes the world.
• Decision. Agents that each give a view within their own specialism, and a meta-orchestrator that weighs those views into a single position decision.
• Execution. The layer that manages the risk budget, slippage and staged entries. A right decision badly executed still loses money, so execution is treated as a specialism in its own right.

We keep performance figures and internals private. But there is a great deal we can say about the architecture. A word of caution, too: nothing here is investment advice. Automated trading carries real risk of loss, and no architecture removes it — our guide on what to check before commissioning a trading bot sets those risks out plainly.

What we learnt

1. The kill switch is the first line of the architecture. Being able to take the system offline in one click is not a feature to add later. In CAI, manual override always wins over automatic decisions. Building that in from the start was the precondition for having the confidence to go live at all.

2. The narrower an agent’s task, the better. A broad brief such as “analyse the market” produces inconsistent results. An agent that answers one question — “trend or compression?” — is both more accurate and testable.

3. The orchestrator does not average the agents. Combining views with equal weight gives the worst agent as much say as the best. Knowing which agent to trust, and how much, under which conditions — that is the real intelligence of the system.

4. Every decision should be logged with its reasoning. When something goes wrong, the first question is “why?”. Without a record of the views that formed the decision, there is no answer.

5. Risk and cost budgets are written first. How much loss is permitted, how many resources an agent may consume and when human approval is mandatory should be set at the design stage — not afterwards.

Where does it help beyond trading?

The same architectural approach applies to decision processes well outside trading — indeed, it has since seeded private consulting work in defence and fintech:

• Systems that pre-assess applications, requests or documents and present a decision to a person, with its reasoning.
• Assistants that work with internal documents and show which information they relied on.
• Monitoring systems that combine data from several sources, report on it and flag anything suspicious.

The common thread: the AI does not decide alone. It produces the parts of a decision, combines them and presents them with reasons. The final word stays with a person, at the point you choose in the design.

Frequently asked questions

Is a multi-agent system always better than a single model?

No. It brings more parts and more connections to maintain. If your task does not combine different kinds of information, change with conditions, carry a high cost of error or need explaining, a single well-designed model call is usually the better choice.

Do all the agents have to be language models?

No. An agent can be a language model, a statistical model or a handful of plain rules. What makes it an agent is a narrow task, defined inputs and a defined output — not the technology behind it.

Can you share CAI’s results or code?

We keep performance figures and internals private. We are happy to discuss the architecture in depth; anything beyond that is a matter for a direct conversation.

How do you keep a multi-agent system under control?

By designing control in from the first day: a kill switch that overrides everything, human approval at critical steps, spending and risk limits per agent, and a log of every decision with its reasoning.

If you are thinking of building a decision process with AI, or making an existing AI project sturdier, we usually begin with a short architecture study: which agents are needed, what data they will use, where the limits lie and how success will be measured. That study is often the most valuable output of the whole project. See our AI systems architecture service, or talk to us.