RavenDB's multi-agent feature lets developers compose multiple AI agents with clear boundaries rather than building a monolithic orchestration framework. Each agent has its own system prompt, LLM connection, scoped tools, and isolated conversation history stored as documents. A parent agent delegates to specialist subagents via a simple SubAgents configuration entry — no custom router or dispatch logic required. RavenDB handles conversation state, memory sharing, and invocation automatically. The approach mirrors classic software design principles: small, independent components with limited scope. Multi-agents are recommended only when teams need independent development, different models per task, or explicit context isolation — not as a default architecture.
