multi-agent-patterns

Multi-agent coordination distributes reasoning across separate context windows to handle complex tasks. You need this approach when a single agent's context window overflows and tasks require parallel exploration with specialized sub-agents.

Decompose your objective into subtasks and assign each to a specialized agent. Implement a supervisor to route tasks, collect results, and aggregate outputs into a final deliverable while maintaining explicit handoffs and context isolation.

Supervisor/Orchestrator, Peer-to-Peer/Swarm, and Hierarchical designs balance control and parallel exploration. These patterns enforce explicit handoffs and robust coordination protocols to maximize throughput while preserving output fidelity.

Coordination primitives like handoffs, voting, debate protocols, and checkpointing align outputs in a multi-agent swarm by enforcing explicit transfers between isolated agents. They ensure reliable synchronization and prevent context bleed across separate roles.

Avoid a multi-agent approach if your task fits within a single context window or lacks specialized sub-tasks. Multi-agent coordination adds overhead through handoffs and context isolation, making it inefficient for simple, linear objectives.