B1 · Prompt chaining

Principle

The task is split into fixed steps. Each model call processes the output of the previous one, and code can check the result between two steps before going on. The path is written in advance: only the content of each step is left to the model.

The problem it solves

A single prompt that asks for everything at once gives uneven results that are hard to fix: you cannot tell which part failed.

When to use it

• The task breaks down cleanly into fixed subtasks: data transformed in several passes, drafting then proofreading then formatting.
• You want to inspect and test each step in isolation. It is also the best starting point: debug the chain before adding complexity.

When to avoid it

• The steps are independent and can run at the same time: prefer parallelization (B3).
• You need to go back or choose what comes next based on what is found: prefer a state graph (B6) or the agent loop (C1).
• A single call already handles the few steps: stay with an augmented LLM (A1).

Synonyms by vendor

• Prompt chaining (Anthropic, AWS, A. Gulli)
• Sequential / pipeline / linear delegation (Microsoft)
• Sequential pipeline (« la chaîne de montage ») (Google ADK)
• Multi-agent sequential (Google Cloud)
• Sequential process (CrewAI)
• Deterministic chain (Databricks)

Neighbouring patterns

• B2 · Routing
• B3 · Parallelization
• B6 · Custom state graph
• A3 · Structured output

Sources

• Building effective agents — Anthropic (E. Schluntz, B. Zhang), 2024-12-19
• Developer’s guide to multi-agent patterns in ADK — Google Developers Blog (S. Saboo), 2025-12-16
• Agentless: Demystifying LLM-based Software Engineering Agents — C. Xia, Y. Deng, S. Dunn, L. Zhang, 2024