Agent Workflows

Introduction

An agent workflow combines multiple agents, tools, and memory to solve complex tasks. LangGraph models them as state machines; CrewAI models them as teams.

Beginner analogy: Like designing a workflow in Notion, but each task can call an LLM and take action.

Understanding the topic

Core concepts to understand:

• Sequential, parallel, conditional flows.
• State shared between steps.
• Checkpoints for resume/retry.
• Human-in-the-loop nodes for approval.

Syntax reference

Visual workflow / architecture:

Start
  │
  ▼
[Plan] (LLM)
  │
  ▼
[Search]──parallel──[Fetch DB]
       │            │
       └─────┬──────┘
             ▼
        [Synthesize] (LLM)
             │
             ▼
       [Human approve?]
             │
             ▼
          [Execute]
             │
             ▼
            End

Real-world use

LangGraph powers many production agents (e.g. Replit Agent). CrewAI is popular for multi-agent simulations.

Best practices

• Model workflows as graphs — easier to reason about than nested loops.
• Persist state for resumability.
• Always add observability per node.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. How do you model agent workflows?
• Q2. Why use state machines?
• Q3. When do you need human-in-the-loop?