Start with a single agent and tools. Split into multiple agents when roles need different instructions, tools or permissions, when contexts would overflow one prompt, or when parallel specialists speed up the work.
Common patterns: an orchestrator calling specialist agents as tools (it keeps ownership of the final answer and shared guardrails) or handoffs where control passes to a specialist.
Costs: more tokens, harder debugging and compounding errors, so add tracing under one request ID and evaluate end-to-end success.
Loading…