Multiple agents, each with a job.
One agent handling everything hits limits: the context window fills up, the tool set gets unwieldy, the task needs genuinely different expertise at different stages. Multi-agent systems split the work. Each agent gets its own tools, instructions, and scope.
An orchestrator routes tasks between them, collects results, and decides what’s next. The tradeoff is coordination overhead. For simple tasks, one well-equipped agent is usually enough. Multi-agent pays off when the complexity is real: heavy parallelization, diverse tool sets, or information that won’t fit in a single context.