Definition
In a multi-agent system a lead agent (the orchestrator) splits a task, hands parts to specialised sub-agents that may run in parallel, and combines their results. Each sub-agent works with a narrower context and toolset, which can improve quality on broad tasks such as research.
The gains come with costs. Anthropic reported that its multi-agent research system used many times more tokens than a single chat, and that coordination, error handling and evaluation get harder as agents multiply. Orchestration pays off when the task is valuable and splits naturally into independent parts.