Multi-Step Workflows

Introduction

A multi-step workflow breaks one big LLM call into many smaller, focused calls. Cheaper, more accurate, easier to evaluate.

Beginner analogy: Like an assembly line — each station does one thing well. Beats one worker building a car alone.

Understanding the topic

Core concepts to understand:

• Split into: extract → classify → summarise → format → validate.
• Each step has its own prompt, eval, and retry logic.
• Cheaper smaller models for early steps; frontier model only where needed.
• Easier to debug — find which step broke.

Syntax reference

Visual workflow / architecture:

User input
   ▼
[Step 1: extract entities]
   ▼
[Step 2: classify intent]
   ▼
[Step 3: retrieve context]
   ▼
[Step 4: generate answer]
   ▼
[Step 5: validate & format]
   ▼
Output

Real-world use

Every production AI feature with > 1 user-visible action is a multi-step workflow. Frameworks: LangChain, LangGraph, LlamaIndex, DSPy.

Best practices

• One responsibility per step.
• Cheapest model that passes the step's eval.
• Log every step's input/output for debugging.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. Why split into steps?
• Q2. Tradeoffs: more steps = more latency. Mitigations?
• Q3. How do you eval a multi-step workflow?
• Q4. What's the failure mode?