ISOM 260 · Final Project

Build Your Agent

Pick a real problem. Build one agent that solves it. Submit by May 6 at 11:59 PM.

Format
Solo or group of 3
Deliverable
Working notebook + 1-page brief
Due
Wed May 6 · 11:59 PM
Submission
Canvas (async, no live demo)

The brief, in one paragraph

Pick one specific real problem that you or someone you know actually has. Build one agent that solves it — with one primary tool, one core flow, and one meaningful guardrail. Try it on a few different inputs to see how it handles different situations. Submit your notebook and a 1-page brief on Canvas by May 6.

01 // KEEP IT SIMPLE

The "ONE" Rule

Five 'ONEs' to keep this doable. Each one is permission to scope down, not a requirement to do more.

One
Real problem

A specific problem affecting a real person — yourself counts.

One
Real user

Yourself counts. A friend, family member, or club you\'re in counts. Real person, reachable for a quick check.

One
Primary tool

Calculator, search, file reader, API, RAG — pick one focused capability.

One
Core flow

Input → reasoning → output. Keep it linear — multi-agent and branching are not needed.

One
Guardrail

Confidence threshold, approval gate, input/output filter — pick one that fits.

Simple is the goal. One tool is enough. One flow is enough. A small agent that solves one real problem really well is exactly what we\'re looking for — that\'s the whole assignment, not a starting point you have to grow beyond.

02 // DELIVERABLES

What to Submit

Two things, one Canvas submission, due May 6 at 11:59 PM.

Working agent

A Colab notebook (or other deployment) where the agent runs end-to-end on real input. Include your tool, your guardrail, and your test cases.

1-page brief

A short structured reflection — five labeled prompts shown below. Markdown or PDF, one page.

The 1-page brief, structured

Five labeled prompts. About a paragraph each. The point is honest articulation, not polished prose — answer plainly.

  • Problem
    In 1–2 sentences: what specific real problem does your agent solve? Who has it?
  • Real user
    Who is the actual person this is for (yourself counts)? What did they say when you tested it?
  • Your agent
    What tool does it use, what's the core flow, what guardrail did you add and why?
  • Test results
    Which test cases passed, which failed, and what failure modes did you discover?
  • Reflection
    What worked, what didn't, and what would you build with another week?
03 // PROBLEM IDEAS

Eight Categories to Adapt

Not prescriptive — pick one as a starting point or invent your own. The constraint that matters: real problem, real user.

Personal Productivity
  • An agent that reads your week's calendar and proposes how to spend Sunday afternoon based on your stated goals
  • An agent that takes your week's git commits or notes and drafts your Monday standup summary
Study Companion
  • An agent that takes a lecture transcript and produces flashcards on the hardest concepts
  • An agent that grades your practice essay against a rubric and highlights weak paragraphs
Decision Support
  • An agent that helps you evaluate apartment listings against your priorities and flags red flags
  • An agent that compares three job offers on the dimensions you care about (commute, comp, growth, culture)
Research Helper
  • An agent that takes a research question and finds three credible recent sources with attribution
  • An agent that fact-checks specific claims in a paper against its cited references
Job Search
  • An agent that compares a job description against your resume and identifies the gaps to address
  • An agent that drafts personalized outreach to alumni based on their LinkedIn role and your shared interests
Real-World Business
  • An agent for a parent's small business — inventory check, customer reply drafting, simple bookkeeping
  • An agent for a club, RSO, or org you're in — meeting summaries, member onboarding, event coordination
Creative Work
  • An agent that takes a paragraph and offers three rewrites in different tones (formal, casual, urgent)
  • An agent that helps you brainstorm titles, intros, or hooks for an essay or pitch
Repetitive Side Project
  • An agent that handles one repetitive task in something you already do — Etsy shop, freelance writing, tutoring
  • An agent that drafts thank-you notes after networking events using the contact details you captured
04 // STARTER RESOURCES

Where to Begin

If you're extending earlier homework, you have most of this already. If you're starting fresh, these sessions have what you need.

Session 4 — Tool use & agent loop
The foundational `run_agent()` pattern. Calculator, stock, company info tools. Your starter notebook.
Session 5 — Real APIs & memory
Wikipedia API, web fetching, error handling, conversation history. The real-world version of Session 4.
Session 6 — ReAct & multi-agent
Reason → Act → Observe. Multi-agent patterns. The Automation Matrix for deciding what to automate.
Session 7 — Knowledge agents (RAG)
Use this if your agent needs to ground answers in documents you provide.
Session 10 — Guardrails
Approval gates, confidence thresholds, input/output guardrails. Pick ONE for your agent.
Session 11 — Testing & evaluation
How to write 3+ test cases that actually expose problems, not just confirm the happy path.
05 // FAQ

Common Questions

I didn't engage with earlier homeworks. Can I still do this?
Yes. Start with the simplest possible version: ONE tool, ONE flow, ONE real user (yourself counts). The Session 4 starter notebook gives you the agent loop pre-built. Solving an artificially small problem really well beats failing to ship a complex one.
How do I pick a "real problem"?
It has to affect a real person. Yourself counts. A friend, a parent, a club, an org you're in counts. A hypothetical "small business owner" doesn't. The exercise of finding a real user is half the learning. If you're stuck, pick from the suggested categories above and adapt one to your life.
What counts as "ONE primary tool"?
Anything that does one specific job: calculator, web search, file reader, Wikipedia API, an LLM-as-judge, a date/time function, a regex extractor, a database lookup. The agent itself can have multiple steps in its loop, but the *tool* it calls should be one focused capability.
What counts as a guardrail?
Any layer that catches a specific kind of failure before it reaches the user or the world. Examples: confidence threshold ("if certainty < 0.7, escalate to human"), approval gate ("ask before sending email"), input filter ("reject if prompt looks adversarial"), output filter ("don't share PII"). Pick ONE that fits your agent's actual risk surface.
How long should I spend on this?
However much time fits your schedule. The point isn't to hit a specific number of hours — it's to ship a small working agent that solves one real problem. If you're extending an earlier homework, it's shorter. In a group of 3 with split work, it's shorter per person. Trust your judgment, and remember that simple-and-finished is the goal.
What if my agent doesn't work perfectly?
That's completely fine — and often more useful than pretending it does. Note in your brief's Test Results section what worked, what didn't, and what you noticed. Understanding where your agent struggles is part of the learning. You don't need a flawless agent to get full credit; you need a real one that you understand.
Is this the Final Project for the course?
Yes — this Build Your Agent assignment is the Final Project for ISOM 260 this semester. It synthesizes what you've learned across the course: tools, real APIs, ReAct, RAG (optional), guardrails, and testing — applied to a problem you actually care about. Submission-only, graded asynchronously, due May 6 at 11:59 PM.

You\'ve got this.

The point of this project isn\'t to impress anyone — it\'s to leave the course with one working agent you built, that solves a real problem for a real user, that you understand end-to-end. Small and working is the goal. Whatever you ship, you\'ll have something to show for the semester and a foundation you can build on later.

Due Wed May 6 · 11:59 PM · Canvas