Week 13: AI Agents — What They Are and How They Work

Welcome to Module 2. You've built neural nets, trained GPT, called APIs, and built a RAG system. Now you're going to build something that actually does things on its own: an AI agent.

---

What is an AI Agent?

An AI agent is just three things in a loop:

1. An LLM (like Claude) that reads a situation and decides what to do
2. Tools that the LLM can call (search, calculate, read files, etc.)
3. A loop that keeps going until the LLM says "I'm done"

That's it. The LLM thinks, picks a tool, gets the result, thinks again, picks another tool (or answers). This loop is what makes it an "agent" instead of a one-shot chatbot.

---

The ReAct Pattern

The most common agent pattern is called ReAct (Reasoning + Acting). It works like this:

1. Thought: The LLM reasons about what to do next
2. Action: The LLM calls a tool
3. Observation: The tool returns a result
4. Repeat until the LLM has enough info to answer

Read this blog post from Anthropic — it's the best explanation of agent patterns out there:

Building Effective Agents — Read at least the first half. You'll come back to it later.

---

Setup

You need the Anthropic Python SDK. If you don't have it yet:

pip install anthropic

Create a .env file in this folder with your API key:

cp .env.example .env

Then open .env and paste your key:

ANTHROPIC_API_KEY=sk-ant-your-key-here

Don't have an API key? Go to console.anthropic.com, sign in (your parent set this up), go to API Keys, and create one. Keep it secret — never commit it to git.

---

Run the Starter Code

python react_loop.py

This runs a minimal ReAct agent with 3 simple tools. Watch what happens:

• You type a question
• Claude thinks about which tool to use
• The tool runs and returns a result
• Claude thinks again
• Eventually Claude gives you a final answer

Try these questions:

• "What is 42 * 67 + 13?"
• "What is the date today?"
• "Tell me about the Python programming language"

---

Tasks

1. Run the agent with the questions above. Watch the tool calls in the output.
2. Read the code. Understand the loop. Where does Claude decide to use a tool vs. give a final answer?
3. Add a new tool. Ideas: get_random_number, reverse_string, or count_words. Follow the pattern of the existing tools.
4. Try to break it. Ask something none of the tools can help with. What happens?
5. Read the Anthropic blog post linked above. Write one thing you learned in your reflection.

---

Deliverables

• react_loop.py with at least one new tool you added
• reflection.md filled in
• Screenshot or copy-paste of one interesting agent run

---

Stuck?

Copy your error message and paste it into Claude or ChatGPT:

"I'm building a ReAct agent with the Anthropic Python SDK. I got this error: [paste error]. Here's my code: [paste code]. How do I fix it?"

This is the same pattern from Module 1 — use AI to debug AI.