Multi-agent systems · Part of the AI agents guide
Multi-agent systems: when one AI agent is not enough
A multi-agent system is a team of AI agents that share one job. Each agent has a role. One plans, one does the work, one checks it. You use a multi-agent system when a job is too big or too mixed for one agent. Here's how they work and when to build one.
How a multi-agent system works
Think of a small team at work. Nobody does everything. Each person has a job, and someone keeps the work moving. A multi agent system is the same idea, with AI agents.
A lead agent
Takes the goal, splits it into tasks and hands each task to the right agent.
Worker agents
Each one does a single kind of task well: find the data, write the draft, fill the record.
A checker
Reviews the work against your rules before it goes out. Sends it back if it's wrong.
Shared notes
One place where the agents record what's done, so nothing is repeated or lost in a hand-off.
One agent or a multi-agent system?
Start with one agent. Most jobs never need more. Add agents only when you hit one of these.
The job has very different parts
Research, writing and checking need different instructions and tools. One agent juggling all three gets worse at each.
The work needs a second look
A separate checker catches mistakes the first agent can't see in its own work.
Parts can run at the same time
Three agents working side by side finish sooner than one agent doing each part in turn.
Common multi-agent patterns
Manager and workers
One lead agent hands out tasks and pulls the results together. The most common setup.
The line
Each agent does its step and passes the work on, like a production line. Simple to follow and test.
Maker and checker
One agent does the work. A second one reviews it and sends it back until it passes.
The router
One agent reads each request and sends it to the right specialist. Good for support and intake.
What goes wrong with multi-agent systems
More agents means more moving parts. These four problems turn up in almost every build, so we plan for them from day one.
The cost climbs
Every agent uses AI on every run. We work out the cost per job before the build.
Agents go in circles
Two agents can pass work back and forth forever. Each one needs a stop rule.
Details get dropped
Things get lost between agents, the same way they do between people. Shared notes fix most of it.
It's hard to see what broke
Without logs you can't tell which agent got it wrong. We log every step from the start.
How we build a multi-agent system in Australia
We're an Australian team with 200+ apps shipped since 2015. We build the system into your product, and you own all of it.
- 01
Product call
Thirty minutes on your product and the job. We tell you straight if one agent would do.
- 02
Map the roles
We split the job into roles, set the hand-offs and write the stop rules.
- 03
Build one agent at a time
Each agent is built and tested alone first. Then we join them up and test the whole team.
- 04
Launch and watch
It goes live with full logs. We watch the cost and the results, and tune as use grows.
Start here
The free AI product call
One call to work out if your product needs a multi-agent system, or something simpler.
- Whether one agent would do the job
- The roles and hand-offs, if you need a team of agents
- The AI running cost per job, so your pricing holds up
- A fixed quote for the build after the call
- Product call
- 30 minutes, free
- Build price
- Fixed quote
- Running cost
- Shown upfront
Want a plan before a build? The Software Plan Workshop is $5,000, credited to a build.
★ Keep reading
More from the AI agents guide
AI agents
The full guide: what agents are, the main types and every page we've written.
AI agents guide →AI agent development
How we add an agent, or a team of them, to your product.
AI agent development →Agentic AI
The idea behind agents: AI that works towards a goal by itself.
Agentic AI guide →The best AI agent platforms
LangGraph, CrewAI, AutoGen and more: the tools people use to run a team of agents.
Compare the platforms →Multi-agent systems explained
Our earlier post, with a worked example of agents sharing one job.
Read the post →Agent memory design
How agents keep track of what they've done and what they know.
Read the post →Multi-agent systems: questions, answered
What is a multi-agent system?
A multi-agent system is a team of AI agents that share one job. Each agent has its own role and tools. One plans, others do the work, and one checks it. Together they finish jobs that are too big or too mixed for a single agent.
When do I need a multi-agent system and not one agent?
When the job has very different parts, when the work needs a second look, or when parts can run at the same time. If none of those is true, one agent is cheaper and easier to trust.
What is an example of a multi-agent system?
A client report. One agent pulls the numbers. A second writes the summary. A third checks the figures match and the tone is right. A lead agent runs the order and hands the report to a person to send.
How much does a multi-agent system cost?
A custom build is a fixed quote after the product call. More agents means a higher running cost, so we work out the cost per job before we start. If you want a plan first, the Software Plan Workshop is $5,000, credited to a build.
Which platform should I use for a multi-agent system?
It depends on the job and your team. LangGraph, CrewAI and AutoGen are common picks. Our post on the best AI agent platforms compares them side by side. We choose the platform after we map the job, not before.
Do you build multi-agent systems for sales or lead generation?
No. We build multi-agent systems into software products. Sales agents and voice agents for lead generation are handled by a sister brand. Tell us what you need on a call and we'll point you to the right team.
Got a job too big for one agent?
Book a free 30-minute product call. We'll tell you if you need a team of agents or just one, and what it would cost.