Splitting work across multiple model calls with explicit handoffs—each step has a narrower job, cleaner inputs, and easier testing.
Prompt chaining composes smaller prompts: extract → classify → transform → verify, often with code between steps. It is a design pattern for reliability, not a single magic prompt.
Relation to agents
Agents may contain chains and branches; chains need not be autonomous. Combine chains with function calling, validation, and logging via LLM observability.