AI Trailblazers · Week 2 of 5

Building your first
AI agent

Anatomy of an agent · ReAct · Kaggle · build a working DataFetcher.

Shape: 30 min theory + 90 min practical · You'll need: the Week 2 Colab notebook + a free Gemini key

The shape of today

Theory, then a real build

2 halves

Theory (30 min): what makes something an agent + the Kaggle world.
Practical (90 min): build a DataFetcher agent, then touch real data.

Plus a 12–15 min energizer where you become the agent loop.

Part 1 · Theory

Anatomy of an AI agent

Concept · Four parts

Plan → Tools → Memory → Action

🧠 Planning

Break the goal into steps; decide what's next.

🛠️ Tools

Abilities it lacks: search, calculator, code, files.

📓 Memory

Hold what it found so steps build on each other.

⚡ Action

Call a tool, read the result, loop till done.

🕵️ Analogy: a detective

Theory (plan) → check records (tools) → case file (memory) → follow leads (action), looping until solved.

Concept · The loop

ReAct = Reason + Act

Thought   → "I need today's AI-education headlines."
Action   → search_web("AI in education 2026")
Observation → [3 articles]
Thought   → "Now summarize into 3 bullets."
Action   → summarize_text(articles)

🔑 Remember

Reason → Act → Observe → repeat. Every agent you build is a version of this loop.

Concept · Kaggle

Kaggle = your portfolio home

Competitions — real problems + leaderboards ("Getting Started" = beginner).
Datasets — thousands, free, one line to load.
Notebooks — publish your work as a shareable link.

💼 Why it matters

A public Kaggle notebook is a résumé/college-app link that proves you built something real. You'll publish one in Week 5.

⚡ Energizer · 12–15 min · on your feet

The Human ReAct Loop

Teams of 4: Brain (Thought+Action), Hands (tool), Memory (writes it), Narrator.
Round 1: answer with no tools → guess wrong. Round 2: loop with tool cards → find out.

Printable card in Coach HQ → Week 2 energizer

Part 2 · Practical

Build the DataFetcher

Build · Step 1

Tools are just functions

def search_web(query):     # returns mock news (reliable)
    return MOCK_NEWS

def summarize_text(text):   # asks Gemini for 3 bullets
    return gemini(f"Summarize in 3 bullets: {text}")

Note

Mock data keeps class reliable — swap in a real news API anytime.

Build · Step 2 — the agent

The loop is the agent

for step in range(6):        # safety cap
    resp = gemini(history, tools=[...])
    if resp.wants_tool:            # Thought → Action
        out = run(resp.tool, resp.args)
        history += observation(out) # → Memory
    else:
        print(resp.text); break     # goal reached

Build · Step 3

Point it at real data

import pandas as pd
df = pd.read_csv(TITANIC_URL)     # one line
df.groupby("Sex")["Age"].mean()  # avg age by gender

🔎 Preview

Next week you'll wrap queries like this in tools so an agent answers data questions on its own.

Week 2 · Wrap

You built an agent 🎉

Agent = Plan → Tools → Memory → Action, run as ReAct.
You gave an LLM tools and ran the loop.
You loaded real data and asked a question.

🏠 Homework

Explore one "Getting Started" Kaggle competition · jot a first project idea in the shared template.

Next week → Data-analysis agents (tools over the Titanic) + scope your own project.