Multiple AI agents together: the lead agent and its specialists
An agent that does everything ends up doing everything halfway. This course explains how to break down an oversized mission between a lead agent and specialized agents, how they hand off work to each other, where you keep an eye on things, and why it's better to wait for the first agent to run smoothly before adding more.
- You break down a mission when it mixes tasks that have neither the same documents, nor the same rules, nor the same pace.
- The lead agent does not produce: it sorts, assigns, verifies and presents the result to you.
- Each handoff between two agents follows a fixed template: what comes in, what goes out, what blocks.
- You watch at the points where an error costs money, not everywhere.
- One agent first: you only coordinate well what you've already seen work.
A restaurant kitchen runs with a brigade: a head chef at the pass, commis at their stations. Nobody does everything, and the chef doesn't cook each plate: they distribute, taste and send back what doesn't work. AI agent teams work on the same principle. We call it orchestration, or a multi-agent system. This course is for those who already have an agent in service and are wondering how to go further.
IWhen a single mission becomes too large
A mission becomes too large when it mixes different trades. Let's take a carpentry shop. At first, the agent responds to requests coming in through the website form. Then you add quotes, then invoice reminders, then questions about installation timelines. Its instructions grow longer, its documents pile up, and its results become irregular.
Check what applies to you: the list remembers your answers on this device.
If you recognize only one of these signs, a better-written mission is often enough. Splitting has a cost, which we will see below.
IIThe principle: a manager who distributes, specialists who produce
An agent whose mission is to sort the work, assign it to the right specialist and verify what comes back.
It is also called an orchestrator. It does not write the quote or the follow-up. It knows how to read a request, recognize what type it is, and check that a result respects the rules before presenting it to you.
Specialized agents, on the other hand, have a short and precise mission. The quote agent knows the rates and deadlines. The follow-up agent knows overdue invoices and the tone to use. The response agent knows the FAQ and opening hours. Each one only has their own documents in front of them, which reduces confusion and the number of tokens read at each handoff.
IIIHow agents hand off work to each other
The weak point of a team, human or not, is the handoff. An agent that receives an incomplete request completes it by guessing. The rule is therefore simple: each handoff follows a handoff sheet fixed, written once and for all. It says what the specialist receives, what it must deliver, in what form, and what it does if it is missing information: it returns the request to the manager with the question, instead of making something up.
- 1The manager reads and sorts
Monday morning, a request arrives: "I need a quote for an under-stairs closet, and by the way, has my March invoice been paid?" The manager sees two distinct topics and separates them.
- 2It assigns each piece
The first goes to the quote agent with the measurements and attached photos. The second goes to the follow-up agent with the invoice number. Each routing slip specifies the customer, the question, and the expected format.
- 3The specialists produce
The quote agent prepares a priced draft according to the rate grid. The follow-up agent notes that the invoice has been paid and prepares a confirmation sentence. Neither of them sends anything.
- 4The manager checks and assembles
It checks the draft: rates from the grid, possible deadline, no unplanned discount. It assembles a single coherent response, in the company's tone, and flags anything that seems doubtful.
- 5You confirm
The carpenter reads a complete response, corrects an installation deadline, and approves. The customer receives a single message that answers both questions.
IVWhere the human watches
The more agents there are, the greater the temptation to let everything run. You need to do the opposite at first. Place your attention where an error is costly: anything that commits to a price, anything that touches an unhappy customer, anything that goes outside. Let the rest through once it has proven itself.
At ABYS, you validate the first sends manually; after three validations, you can switch to automatic validation, and the risks of this choice are spelled out in black and white. In a team, think agent by agent: the response agent may deserve its autonomy well before the quote agent.
VWhy to start with a single agent
A team of agents is harder to tune than a single agent. When the result is wrong, you need to know who made the mistake: the manager who sorted poorly, the specialist who produced poorly, or the handoff sheet that forgot information. Without experience of a first agent, you cannot read these errors.
A single agent teaches you the essentials: write a clear mission, choose the right documents, review a blank test. When it has been running for a few weeks, the limitations appear on their own, and they tell you which specialist to create. The first split is often obvious: the task you correct the most becomes a separate agent.
VIWhat a team of agents costs
Each agent consumes tokens, and the manager does too: sorting and verification require reading. A team therefore consumes more than a single agent for the same request, in exchange for more reliable work. At ABYS, multiple agents can work starting from the plan at 19 euros excl. VAT per month: they all draw from the same token reservoir, 100,000 tokens per month for this plan, and the account displays consumption. It is this reservoir, not the number of agents, that limits the work. If the team grows, the following plans offer 700,000 tokens at $99 excl. VAT, 1.8 million at $199 excl. VAT and 5 million at $349 excl. VAT.
A carpenter corrects the same kind of error every week, always on invoice follow-ups, while responses to requests are correct. What does this sign suggest?
?Questions we're asked
What is a multi-agent system in AI?
This is an organization where several AI agents work on the same mission, each with its own role. Most often, a manager agent receives the work, breaks it down, assigns it to specialized agents, then checks their results. This distribution makes each agent more precise, at the cost of more delicate tuning and slightly higher consumption.
When should you use multiple AI agents instead of just one?
When a mission mixes tasks that do not have the same documents, the same rules or the same pace, and you often correct the same errors. If your first agent is not yet working well, adding agents will not fix anything: improve its mission and documents first.
How do AI agents communicate with each other?
They transmit work to each other in a form agreed in advance: a sheet that specifies what the agent receives, what it must deliver and how it signals missing information. The stricter this sheet, the less agents have to guess. Each handoff must leave a trace in the log so you can track the request.
Do multiple AI agents cost more?
They consume more tokens for the same request, because sorting and checking require reading. At ABYS, the price doesn't depend on the number of agents: from 19 euros excl. VAT per month, all agents in the account share the same token pool, and it's this pool that determines the amount of work possible.
Start with one agent, on a single well-defined mission. When it runs well, its limitations will show you which specialist to add.
Create My AgentRead more next
Writing an AI agent's mission: the five-point method
An agent works as well as their mission is clear. This course gives you a five-point method to write it: the goal, the…
Read the lessonHave the final say on an AI agent: validate, then delegate
An AI agent prepares the work, you decide what goes out. You'll know at the end how to review their first sends, what signs…
Read the lessonAI tokens: what an agent consumes and how to predict it
Tokens are the unit of work for AI: everything an agent reads and everything it writes is counted in tokens. After this course, you…
Read the lesson