AI Workflow Design

Introduction

An AI workflow is a multi-step pipeline that combines LLM calls, retrieval, tool calls, and conditional logic to solve a task. Think of it as a flowchart where some nodes are AI calls.

Beginner analogy: Like a Zapier zap, but each step might call an LLM.

Understanding the topic

Core concepts to understand:

• Break complex tasks into smaller LLM calls.
• Routers — first LLM call decides which branch to take.
• Parallelisation — run independent steps concurrently.
• Reflection — LLM critiques its own output and retries.
• Tools: LangChain, LangGraph, Vercel AI SDK.

Syntax reference

Visual workflow / architecture:

       Input
         │
         ▼
   ┌──────────┐
   │ Router   │  classify intent
   └─┬─────┬──┘
     │     │
     ▼     ▼
  [Search] [Summarise]
     │     │
     └──┬──┘
        ▼
   ┌──────────┐
   │ Reflect  │  critique + retry
   └─────┬────┘
         ▼
      Output

Real-world use

Perplexity uses a multi-step workflow: query → search → re-rank → summarise → cite. v0.dev uses a workflow: prompt → plan → generate → critique → fix.

Best practices

• Smaller steps = easier to debug and evaluate.
• Parallelise where independent — saves latency.
• Always log every step for postmortems.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. Why split a task into multiple LLM calls?
• Q2. What is a router pattern?
• Q3. What is reflection?