multi-agent-patterns

Multi-agent architectures distribute work across specialized agents to overcome single-agent context limits. This skill covers three patterns—Supervisor/Orchestrator for centralized planning, Peer-to-Peer/Swarm for distributed collaboration, and Hierarchical for layered execution—each with coordination primitives, task decomposition, and failure handling to scale reasoning across domains.

Three core patterns structure multi-agent coordination: Supervisor/Orchestrator centralizes planning and task distribution; Peer-to-Peer/Swarm enables agents to handoff work directly; Hierarchical layered strategy separates planning, coordination, and execution. Each enforces context isolation and messaging protocols for reliable collaboration.

Use multiple agents when tasks decompose into subtasks requiring parallel execution, specialized toolsets, or domain-specific expertise across roles. Multi-agent designs excel when single-agent context limits constrain reasoning, cross-domain capabilities are needed, or independent reasoning streams accelerate complex problem-solving.

Context isolation separates each agent's knowledge and tools by explicit role definition and messaging boundaries. State management tracks task progress, agent availability, and coordination state through centralized or distributed ledgers, with failure handling and retry logic to maintain consistency across handoffs.

Yes, multi-agent architectures enable parallel execution through task decomposition and handoff protocols. Coordination primitives—supervisor routing, peer messaging, or hierarchical delegation—merge agent outputs, resolve conflicts, and synthesize results into coherent task completion.