Multi-agent systems explained: when one agent is not enough
Multi-agent systems share a job between several AI agents and pass work from one to the next. An AI agent is a software helper that reads, thinks and hands back a result. Several agents help when one agent has so many different duties that it does none of them well. For most small-company tasks, one agent is enough.
What it looks like: a reception desk
Think of the front desk in a medical practice. One person greets every visitor. They answer simple questions themselves and send everyone else to the right specialist.
A multi-agent system works the same way. A general agent receives your request. Behind it sit specialists: one knows the customer records, one writes text, one checks goals. The general agent decides who fits, hands the work over and brings back the result. You talk to one helper and never see the hand-over.
Why split the work at all?
- Short instructions work better. One agent with three pages of rules for five jobs mixes them up. Five agents with half a page each do not.
- Less can go wrong. An agent that only writes text does not need the right to change your records.
- You can fix one part. Improve the specialist and the others stay as they were.
The split has a price. Every hand-over takes a little time and costs a little money. Detail can also get lost on the way, as in any message passed along.
One agent or several? Four signs
| Look at | Stay with one agent | Split into several |
|---|---|---|
| Instructions | Fit on one page | Rules for different jobs contradict each other |
| Kinds of work | One, such as summarising | Clearly different, such as looking up data and writing text |
| Rights | The same for the whole job | One part may only read, another may suggest changes |
| Reuse | The job is needed in one place | One part is needed in many places |
Two or more marks in the right-hand column are a reason to split. One mark is not.
A small example
A team lead asks the assistant: "Write a one-page brief for the autumn campaign from our open tasks, with a short headline." The general assistant reads the tasks itself, because that is its normal work. For the headline it asks a small writing agent, which returns one line. The assistant puts both together and shows the page for review.
Two agents, one hand-over, one conversation. That is the scale at which this pays off in a small company: one general helper plus a few narrow specialists.
Three rules that keep it manageable
- Give every agent a one-line job. That line is how the general agent picks the right specialist.
- Let only one agent talk to the person. Specialists report back to the agent that asked.
- Keep rights with the person. No agent in the chain should see more than the person who started it.
How Worknano handles this
In Worknano every agent has a list of other agents it may hand work to. The Generic Assistant Planner is the general assistant behind the sparkles button. It answers most requests itself and hands a question to another agent when a specialist fits better. Page assistants such as the Task Assistant know the habits of their own page. Whichever agent does the work, it reads only what the signed-in user may read, and every change is a preview you confirm.
- Generic Assistant Planner: the assistant behind your page sparkles
- AI agents: automate tasks with prompts, not code
- Build your own agent and set who it may delegate to
Create a free workspace and watch a general assistant hand a question to a specialist.