Common patterns: the supervisor pattern (one "manager" agent delegates to specialized "worker" agents), the debate pattern (agents argue opposing positions to reach a more balanced conclusion), the pipeline pattern (agents process sequentially, each refining the previous output), and the peer pattern (agents work in parallel on different aspects of a problem and merge results).
Several frameworks support multi-agent systems: AutoGen (Microsoft) enables agents to converse with each other, CrewAI provides role-based agent teams, LangGraph handles complex agent workflows as state machines, and Anthropic's agent SDK supports multi-agent orchestration. The choice depends on complexity: simple handoffs don't need a framework; complex workflows with branching logic and human-in-the-loop approval benefit from structured orchestration.
Multi-agent systems multiply LLM API costs — if three agents each make five calls to solve a problem, that's 15x the cost of a single call. The value proposition is that the quality improvement justifies the cost for high-stakes tasks. A code review agent that catches bugs before deployment saves more than the API calls cost. But for simple tasks, a single well-prompted model is usually sufficient and far cheaper.