When is a multi agent system better than a single agent?

Answer

The honest answer is that multi agent systems are used more often than they are justified.

They genuinely help in two cases:

Context isolation. A subagent researches something across thirty files and returns a summary. The thirty files never enter the parent's context. This is the strongest real argument, and it is about managing the context window, not about specialisation.

Genuinely independent parallel work. Several subtasks with no shared state and no ordering dependency can run concurrently for wall clock savings.

Where it hurts:

Coordination cost. Passing information between agents means serialising it to text and re-parsing it. Detail is lost at every hop, and the losses compound.

Error propagation. One agent's confident mistake becomes another's premise, and the second agent has no way to know it was wrong.

Cost and latency multiply. Each agent has its own context and its own model calls.

Debugging becomes very hard. Tracing why the system produced an answer across five agents is materially harder than tracing one.

A well configured single agent with good tools usually beats a poorly coordinated committee. Reach for multiple agents when you can name the specific context or parallelism problem being solved.

Common wrong answer

Assigning agents role labels such as "architect", "coder", and "reviewer" and expecting that to improve quality. Role prompts change tone more than capability. The same model is behind each one. Unless the roles have genuinely different tools or genuinely isolated context, you have added cost and hops without adding ability.