Single vs Multi-Agent Systems: When Splitting an Agent Earns Its Cost

Multi-agent designs split work across several model-driven agents that hand tasks to each other. They add handoffs, each of which can lose context and compound errors, plus latency, cost and debugging difficulty. This guide explains the costs, the four situations where splitting pays (permissions, parallelism, review, separate domains), and the common patterns.