How do you implement multi-agent systems where AI agents collaborate on complex tasks?

Multi-agent systems decompose complex tasks across specialized agents that collaborate. Architectures: (1) Hierarchical: a supervisor agent breaks the task into subtasks and delegates to specialist agents (researcher, coder, reviewer), collecting and synthesizing results. (2) Sequential pipeline: agents process in order, each receiving output from the previous agent (planner creates a plan, executor implements, reviewer validates). (3) Debate/consensus: multiple agents independently solve the same problem, then discuss and vote on the best solution. (4) Swarm: agents dynamically self-organize based on task requirements without central coordination. Implementation with frameworks: CrewAI (role-based agent teams with configurable collaboration), AutoGen (Microsoft's multi-agent conversation framework), LangGraph (graph-based agent workflows with state management), and custom implementations using message passing. Agent design: each agent has a specific role (defined via system prompt), available tools (scoped to its responsibility), and communication protocol (structured message format). Communication patterns: shared scratchpad (agents read/write to a common state), message passing (agents send structured messages), and blackboard (agents post findings to a shared knowledge base). Coordination: implement task decomposition with dependency tracking, handle parallel execution of independent subtasks, implement timeout and fallback for unresponsive agents, and maintain conversation history for context continuity. Production considerations: implement cost controls (limit total LLM calls per task), set execution time limits, implement human-in-the-loop checkpoints for critical decisions, and log all agent interactions for debugging and improvement.