AI Agent Exercises

Introduction

Practice cements the theory. This lesson lists 10 hand-picked exercises that take you from 'I read the docs' to 'I shipped an agent'. Do them in order; each builds on the last.

Beginner tip: resist the urge to copy code — type each exercise by hand. The muscle memory is the lesson.

Understanding the topic

Core concepts:

• 1. ReAct agent with a single calculator tool.
• 2. RAG bot over a small markdown folder.
• 3. Plan-Execute agent with a web-search tool.
• 4. Reflection loop on code generation.
• 5. Supervisor + 2 specialists for a research task.
• 6. Long-running agent with persisted state.
• 7. Cost-capped autonomous email triager.
• 8. Tool-call validator wrapping unsafe shell tool.
• 9. Prompt eval harness with 20 cases.
• 10. Production deployment to Vercel + Inngest.

Syntax reference

Visual workflow / architecture:

┌──────────────┐
│   Exercise   │
└──────┬───────┘
       ▼
  Build · Test · Eval
       │
       ▼
  Compare with model answer

Real-world use

These mirror real interview tasks at Anthropic, OpenAI, Cursor, Cognition and enterprise AI teams.

Best practices

• Time-box each exercise (≤ 2 hours).
• Write the eval BEFORE the agent.

Common mistakes

• Skipping evals — you never know if you regressed.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. Which exercise is foundational for all the others?
• Q2. Why write evals first?
• Q3. Scenario: an exercise stalls. What's your debugging order?