What Is a Multi-Agent System? – Definition and How It Works
Learn what a multi-agent system is, the common coordination topologies, what specialisation buys, and why single agents often outperform them.
A multi-agent system uses several model-driven agents that coordinate on a task, each with its own role, tools, and often its own model. The design question is not whether it can work, but whether the division earns its coordination cost.
1. The Topologies
| Topology | Structure | Suits |
|---|---|---|
| Sequential | Output of one feeds the next | Pipelines with a known order |
| Supervisor | A manager delegates and collects | Order that must be decided at run time |
| Parallel | Several work simultaneously, results merged | Independent sub-tasks |
| Debate | Agents critique each other before concluding | High-stakes judgements |
Sequential is the most predictable and the easiest to debug. Supervisor topologies are the most flexible and the hardest to trace, because the plan itself is generated.
2. What Specialisation Actually Buys
The gain comes from scope, not from personas. An agent restricted to one tool and one task makes fewer wrong choices, for the same reason a narrow prompt outperforms a broad one.
Giving an agent an elaborate backstory changes little; giving it two tools instead of twelve changes a great deal.
3. The Costs
- Latency and spend. Every agent adds model calls, and supervisors add planning calls on top.
- Error propagation. A weak early output is elaborated confidently downstream.
- Coordination failure. Agents duplicate work, or each assumes another handled something.
- Debugging difficulty. Traces interleave, and the causal chain is harder to follow.
- Context duplication. The same material is often re-supplied to several agents.
4. When It Genuinely Helps
- Sub-tasks need different tools, and giving one agent all of them would degrade selection.
- Independent work can proceed in parallel with a real wall-clock benefit.
- Review is valuable — a separate checking agent catches what the producing agent missed.
- Different steps warrant different models for cost or capability reasons.
5. When One Agent Is Better
If the task is one loop over one tool set, a single agent is faster, cheaper, and far easier to debug. Splitting a task with no natural seam adds handoffs without reducing the work.
A useful test: could the split be explained to a colleague as a real division of labour? If it sounds like organisational decoration, it probably is.
6. Making Handoffs Reliable
The interface between agents matters more than the agents themselves.
- Define each agent's output as a schema, not as prose.
- Validate between steps, so a bad output fails there rather than three steps later.
- Pass references to large artefacts rather than their contents.
- Cap iterations per agent and across the system.
- Trace every agent's calls with a shared run identifier.
7. Evaluating the Design
Measure against a single-agent baseline on the same tasks. Multi-agent systems frequently look impressive and perform no better, and only a comparison reveals which case you have.
Compare quality, cost, and wall-clock time together — a modest quality gain at several times the cost is a real trade-off, not a win.
8. Summary
A multi-agent system coordinates several agents through sequential, supervisor, parallel, or debate topologies, with the benefit coming from narrowed scope rather than from roles.
It earns its coordination cost when sub-tasks need genuinely different tools or independent review. Otherwise a single well-scoped agent is cheaper, faster, and easier to fix — and the only way to know which you have is to measure both.