Multi Agent Patterns

This walks you through the three main patterns for splitting work across multiple LLM instances: supervisor (centralized control), swarm (peer handoffs), and hierarchical (layered abstraction). The key insight is that you're doing this for context isolation, not role-playing organizational structure. Watch out for the "telephone game" problem where supervisors lose 50% performance by paraphrasing sub-agent responses instead of forwarding them directly. The token economics section is especially honest about costs, showing multi-agent systems can burn tokens much faster than single-agent approaches. Use it when you're actually hitting context limits or need true parallelization, not just because multiple agents sound more sophisticated.