AI Trailblazers · Week 4 of 5

Independent project:
the hackathon

Debug agents like a pro, then build your own MVP in a structured, time-boxed sprint.

Shape: 20 min theory + 100 min practical · You'll need: the Week 4 Colab notebook + your project scaffold

The shape of today

A short clinic, then you build

2 halves

Theory (20 min): how agents break, and how to see the break before it costs you an hour.
Practical (100 min): a phased hackathon — build, check in, expand, debug together.

Short theory on purpose — today is about your hands on your keyboard.

Part 1 · Theory

Debugging agentic systems

Concept · Why agents break

Three failure patterns

🌀 Hallucination

The model invents a fact or a fake tool result instead of calling a tool.

🔧 Tool error

Wrong args, bad URL, empty return — the tool ran but gave garbage.

📝 Prompt issue

Vague instructions → the agent loops, quits early, or ignores its tools.

⚠️ The trap

All three look the same from the outside — "wrong answer." Naming which one is 80% of the fix.

Concept · See the break

Log every Thought → Action → Observation

You can't fix what you can't see. Print each step so the failure point is obvious.

for step in range(6):
    print(f"🧠 Thought: {resp.thought}")
    try:
        out = run(resp.tool, resp.args)
        print(f"🔧 Action: {resp.tool}({resp.args}) → {out}")
    except Exception as e:
        print(f"❌ Tool failed: {e}")   # catch, log, keep going
        out = f"ERROR: {e}"              # feed error back to agent

🔑 Remember

A logged loop turns "it doesn't work" into "it broke on line 3, tool error." Add prints before you ask a coach for help.

⚡ Energizer · 8–10 min · teams

Bug Hunt 🐛

Each team gets a printed agent transcript with 3 planted bugs.
Find each one, classify it: hallucination / tool error / prompt issue.
Write a one-line fix for each. Fastest correct team wins.

Printable transcript in Coach HQ → Week 4 energizer

Part 2 · Practical

The structured hackathon

Practical · The plan

Four phases, two check-ins

Phase 1 · Foundation — 30 min: core loop + first tool + data source.
Check-in — 15 min: share progress, solve one common issue.
Phase 2 · Expansion — 40 min: 2nd tool + error handling + end-to-end.
Debugging Clinic — 15 min: submit blockers, vote, solve together.

🎯 Target

By the end: a working MVP that uses 2 tools and answers a real question. Not polished — working.

Phase 1 · Foundation · 30 min

Get one thing working

Wire the core ReAct loop from the scaffold — don't rebuild it, adapt it.
Add your first tool (one function that returns something real).
Connect one data source (a CSV, an API, or mock data).

✅ Phase-1 done =

Agent runs the loop, calls your one tool, and prints an answer — even a rough one.

Check-in · 15 min · together

Share progress, unblock as one

Each student/team: one sentence on what works + one thing that's stuck.
Coach picks the most common blocker and solves it live for the room.
Quick wins get a shout-out — momentum is contagious.

🗣 Format

Round-robin, 15–20 sec each. No debugging in the circle — just name it. Fixes happen live after.

Phase 2 · Expansion · 40 min

Second tool + make it robust

Add a 2nd tool or data source so the agent can chain steps.
Wrap tool calls in try/except and log — feed errors back to the agent.
Run it end-to-end on a real question and read the full transcript.

def query_data(question):
    try:
        return df.query(question)      # real work
    except Exception as e:
        return f"Couldn't run that: {e}"  # graceful, agent recovers

Debugging Clinic · 15 min · together

Crowd-solve the hard bugs

Everyone drops their #1 remaining blocker in the shared doc.
Room votes — pick the top 2 most-shared or most-interesting.
Solve each live, naming the pattern: hallucination / tool error / prompt issue.

🔑 Why public

One student's bug is usually five students' bug. Solving it out loud teaches the debugging method, not just the fix.

Support · Meet everyone where they are

Stretch or scaffold

🚀 Advanced stretch

Add a simple UI (Gradio), or a multi-agent setup — one plans, one executes.

🧩 Extra support

Pre-built tool templates to drop in, plus pair programming with a coach or peer.

🎚️ Same finish line, different paths

Everyone leaves with a working 2-tool MVP. How far past that is up to them.

Week 4 · Wrap

You built your own agent 🎉

You can name the 3 failure patterns: hallucination · tool error · prompt issue.
You logged the loop and used try/except to see and survive bugs.
You shipped an MVP with 2 tools that answers a real question.

🏠 Homework

Get your MVP to a clean, runnable state — 2 tools working end-to-end. Jot 3 bullets: what it does, one thing you're proud of, one thing to improve.

Next week → Polish & present: publish your Kaggle notebook and demo your agent to the group.