Lehigh University · Agentathon
Build an
AI agent.
90 minutes to learn. 3 hours to ship.
today
20 min
What is an AI agent?
And why do most of them fail silently
50 min
Build one. Live.
System prompt → tools → actions → UI
20 min
Does it actually work?
How to know when your agent breaks
3 hrs
You build. We help.
Fork the starter. Solve a real Lehigh problem.
Quick
show of hands.
Used
ChatGPT?
Used
Perplexity
or Devin?
That's
the difference.
chatbot
answers
questions.
agent
completes
tasks.
the difference at Lehigh
ChatGPT tells you
how to register
for CSE 340.
An agent checks
if it's full.
Finds the next
open section.
Texts you.
what makes agents different
  • →They use tools — search the web, call APIs, read files
  • →They reason — plan multi-step tasks, make decisions
  • →They act — send emails, set reminders, create calendar events
  • →They remember — maintain context across a full conversation
but here's the thing
Most of them
are broken.
we test 6,200+ agents every day · real data
Here's what we found
in March.
4.5M
tests run
56%
perfect uptime
89%
wrong answers
0.2%
fully healthy
The server returned
200 OK.
The answer was
wrong.
Grafana said
all clear.
Users got
wrong answers.
Building agents
is easy.
Knowing if they
work is hard.
real talk
"Why not just
use Claude or
ChatGPT?"
Claude answers
your questions.
An agent answers,
then does something
about it.
the actual difference
  • →Claude waits for you. Agents run while you sleep.
  • →Claude answers. Agents post to Discord, send emails, call APIs.
the actual difference
  • →Claude is a conversation. An agent is a product you ship.
  • →Claude responds to prompts. Agents respond to events.
Part 2 · 50 min
Let's build
one.
We're solving a
real Lehigh problem.
Registration
is painful.
what students actually ask
  • →"Is CSE 340 full? When does a seat usually open?"
  • →"Which section of MATH 231 has the better professor?"
  • →"What are the prereqs and do I have them?"
  • →"If I drop this, what happens to my financial aid?"
  • →"Alert me the second a seat opens."
what we're building
step 1
System prompt
Give the agent its personality and purpose
step 2
Add web search
Agent can now find real Lehigh information
step 3
Add actions
Agent does things — emails, saves, alerts
step 4
Add a UI
One line of Gradio — public URL anyone can use
open your notebook now
6 cells.
Run them
top to bottom.
Go to agentstatus.dev/lehigh → File → Save a copy in Drive first.
Cell 1 · Install
Installing the
dependencies.
Downloads OpenAI, Tavily, and Gradio into your session.
Takes about 30 seconds. Run it once.
Cell 2 · Keys
Loading your
API keys.
Keys live in the 🔑 Secrets tab on the left — not in the code.
Add OPENAI_API_KEY and TAVILY_API_KEY there.
Cell 3 · System prompt
The system prompt
is the soul
of the agent.
Cell 3 · System prompt
It defines personality,
purpose, and limits.
This is the first thing you change during hack time.
"You are a Lehigh registration assistant" → "You are a Lehigh dining scout"
Same code. Completely different agent.
Cell 4 · Tools
Tools are how
it does things.
Not just
says things.
Cell 4 · Tools
Two tools built in:
search and post to Discord.
The agent decides when to call a tool.
You define what happens when it does.
Adding a new tool = one new function + register it.
Cell 5 · Agent loop
The core logic.
Don't touch this.
Sends your message → model decides to search or answer → tool runs → answer comes back.
Every response automatically posts to #agentathon on Discord.
Cell 6 · Launch
One cell.
Public URL.
Anyone can use it.
It's not a chatbot
anymore.
It acts.
Part 3 · 20 min
Does it
actually
work?
OpenAI pushed
an update
last night.
Your code
didn't change.
Your answers
did.
Nobody
told you.
agent status · agentstatus.dev
We watch your agent
from the outside.
UP
Responded. Answer passed all checks.
DEGRADED
Responded. But the answer was wrong.
DOWN
Couldn't reach the agent at all.
AgentStatus · agentstatus.dev/lehigh
500 free
credits.
Every team gets 500 free AgentStatus tests.
No credit card. Just sign up and go.
hack time
3 hours.
Build something
real.
fork the starter · agentstatus.dev/lehigh
Open the notebook. Add your two keys. Run all cells.
Every response posts to #agentathon automatically.

Seat alert agent

Monitor CSE 340. Text you the second a seat opens.

Schedule builder

Takes your requirements, builds a conflict-free schedule.

Professor ranker

Compares every section of a course by real reviews.

Degree auditor

Checks if your current courses put you on track to graduate.

Dining scout

Tells you what's at each dining hall right now.

Your idea

What's the most annoying thing about being a Lehigh student?

The best agents
solve one real problem
for one real person.
judging · demo at the end
  • →Does it work? Live demo, no screenshots
  • →Does it use tools? Has to do something, not just chat
  • →Does it take actions? Sends, saves, or alerts something
  • →Would people use it? Solves a real Lehigh student problem
go
Build something
you're proud of.
Starter: agentstatus.dev/lehigh · Monitor: agentstatus.dev
Questions: dulra@carmel.so · roman@carmel.so
← → arrows · click · F fullscreen