Multi-agent system
A system where several LLM agents with different roles collaborate to solve a task — e.g. a planner, a coder, and a reviewer, or several that debate. The idea of dividing work among specialized agents instead of one all-knowing one.
Architectures: orchestrator-worker (a main agent delegates sub-tasks to specialists and gathers the results), role-based (researcher → writer → editor in a pipeline), debate/collaboration (agents critique each other's answers for a better final result). Arguments for: separation of concerns, each agent can have its own prompt/tools/context, easier to reason about. Arguments against: more complex, more expensive (many calls), and sometimes a single well-equipped agent does the job just as well (Anthropic and others warn against over-engineering). Frameworks: CrewAI, AutoGen, LangGraph. Hot but debated: when do several agents pay off vs one good agent? A central design question in agentic systems. Related to agentic workflow and ReAct agent.