Multi-agent systems split work across specialized roles (planner agent, researcher agent, coder agent, critic agent), each with its own prompt, memory, and tools, coordinated by a supervisor or message bus. Advantages: (1) separation of concerns: clearer prompts; (2) parallelism: independent subtasks run concurrently; (3) self-critique: a critic agent can review another agent's work. Disadvantages: (1) communication overhead; (2) compounded errors: if the supervisor misunderstands, all subagents suffer; (3) resource cost: 2-5x more LLM calls. Empirically (2024-2026), multi-agent helps in coding workflows (Coder + Tester + Reviewer competition) and deep research (Planner + Searcher + Synthesizer). For simple workflows it adds latency and risk without quality. Senior rule: prefer single agent with good tools; reach for multi-agent when (a) tasks need distinct role expertise, or (b) when self-critique materially improves output.