When is multiple agents justified? Return JSON {"when_yes": [list of 2+ cases], "when_no": "when not needed"}.
Multi-agency is fashionable, but most tasks require ONE well-tuned agent. Multiple agents are justified when: the task breaks down into clearly DIFFERENT roles or skills (e.g., researcher plus critic plus writer); parallelism is needed (many independent subtasks at the same time); independent control is useful (one agent checks the other). It is not necessary when: the problem is solved by one prompt or a simple chain of steps; you split for the sake of “cool”, not for the sake of business - each agent adds cost, delay and points of failure. Rule: add agents only when the role or concurrency really requires it; the default is one good agent.