Multi-agent system: definition and use
Why the Term Matters
Complex workflows often fail because a single assistant is asked to do too much at once. A multi-agent system splits the task into roles with clear boundaries, which makes each part checkable. The appeal is also the danger, because more agents mean more moving parts. A prototype honestly shows whether the split really helps or only creates complexity that a good single assistant would not need.
When It Is Relevant
- A task naturally breaks into clearly separated steps.
- Different steps need different tools or knowledge.
- A review step should check another agent's work.
- Long workflows benefit from specialisation rather than one all-rounder.
When Not
If a task is solvable in one step, a single assistant is faster, cheaper and easier to maintain. Several agents raise latency, cost and sources of error. As long as a single agent handles the task cleanly, the split is usually premature and only adds effort without real gain.
Industries That Use It
- Audit and assurance: one agent collects evidence, a second checks against rules, a third summarises, with clearly separated roles.
- Media and publishing: research, fact-check and editing run as separate steps with their own control.
- Construction: review tenders, check quantities and compare bids as bounded subtasks.
- E-commerce: enrich product data, create copy and check against guidelines in a chain of specialised agents.
Typical Use Cases
- A research agent collects sources, a review agent judges how solid they are.
- A planning agent breaks down a task, execution agents handle the parts.
- A writing agent drafts, a control agent checks against rules.
- A coordinator routes requests to the right specialist agent.
- A review agent stops the flow when an intermediate result is not plausible.
Decision Guide
| Signal | Multi-agent system fits | Alternative |
|---|---|---|
| Task breaks into clear roles | specialised agents with a coordinator | single assistant |
| Solvable in one step | rather not | single agent |
| Second-instance review wanted | review agent as its own role | self-check in the prompt |
| Cost and latency critical | keep the setup lean | as few agents as possible |
Example
A publisher wants to prepare research articles. In the workshop a prototype emerges with three roles: an agent collects sources, an agent checks how solid they are, an agent writes the draft. On real topics it becomes clear where the split adds quality and where a single assistant would have sufficed. This describes the method, not an asserted client metric.
Service Connection
The right entry is the AI workshop, then AI development and AI automation.
Sources
This framing describes common practice around multi-agent systems. Whether and how a split makes sense depends on the concrete workflow and should be checked on the real case.