Multi-agent demos are extremely convincing. One agent becomes a researcher, another a planner, another a critic and another a manager. It looks like a tiny software company appearing inside a terminal.
The uncomfortable question is whether the task needed a company.
Every handoff between agents loses something. Context has to be summarized, goals have to remain aligned and contradictory results need reconciliation. If one capable agent can complete the task with good tools and a manageable context window, adding more agents can reduce reliability rather than improve it.
There are still strong reasons to use multiple agents. Parallel research is an obvious example. Several independent searches can explore different branches of a broad topic simultaneously. Context isolation is another. A specialized sub-agent can process a large document or repository and return a compact result without filling the coordinator's working context.
Specialized permissions can also justify separation. An analysis agent may have read-only access while an execution agent has carefully constrained write tools. The architecture then represents a real security boundary instead of a role-playing exercise.
The coordinator matters. A multi-agent system needs an explicit contract for what each worker receives, what it returns and how the overall task determines completion. Without that contract, agents can produce beautifully written summaries of slightly different problems.
Token cost grows quickly as well. Every worker needs instructions and context, and the coordinator then consumes the workers' outputs. Parallelism can improve wall-clock time while still multiplying total inference cost.
I start with a single agent plus deterministic application code. When I hit a measured limitation such as context overload, independent parallel work or genuine specialization, I introduce another agent for that specific reason.
Multi-agent architecture should therefore be an optimization, not an aesthetic. If the explanation for an extra agent is merely "this one is the senior researcher," the system is probably imitating an org chart instead of solving an engineering constraint.