- Class
- Tuesdays, 10:10am–noon
- Location
- 602 Northwest Corner
- Instructor
- Junfeng Yang
- Office
- 519 CSB; office hours by appointment
- Course
- 3 points; section 021; call number 19731
- Contact
- Private note on Ed, not email
Admission is closed for Fall 2026. See the admission archive and commitment and communication policy.
What this course is
You will build and launch an AI agent product, get real users, and try to get one of them to pay. You do it in a team of two or three, on a weekly cadence, in public in front of every other team.
This is not a survey of agent architectures and it is not a business plan course. There is no midterm and no final exam. There is a number you report every week and a room full of people who notice when it does not move.
What it costs you
- Roughly 10 hours a week outside class. More in weeks 6, 7, 12, and 13.
- Your product must be live and used by someone outside the class by week 7. This deadline does not move.
- You report a number publicly every week, including the weeks it is bad.
- You are in the room. Every team submits a weekly update; oral team rounds run during regular sessions. Guest sessions include discussion and team deep dives. See the session formats.
If any of that is a problem, take a different course. You will be happier and so will your teammates.
Teams
- Size is 2 or 3. Not one, not four.
- Teams are fixed at admission. You were admitted as a unit, so you cannot re-form into different teams once the term starts. If your team is genuinely broken, talk to the instructor rather than reshuffling on your own.
- If you lose a member: a team of three continues as two. A team of two that drops to one joins another team as a third. Because admission is by team, there is no bench of individuals to backfill from. This is one more reason your choice of teammates is the most consequential decision you make before the term starts.
- Idea veto. In week 2 the instructor will veto ideas that cannot reach real users in twelve weeks. Regulated health, hardware, and anything with an enterprise sales cycle longer than a term are out. This protects you from spending the term on something structurally unwinnable.
Grading
Weekly deliverables, Demo Day, and the final submission receive a shared team grade. Classroom contribution is graded individually. The instructor assigns all course grades; feedback from guests and judges is advisory.
| Weight | Component |
|---|---|
| 40% | Weekly deliverables. Twelve submissions, weeks 3 through 14, each including the assignment and weekly update. Lowest grade dropped. Graded on what you built and tested, evidence from users, and honest reporting of what you learned. Not on whether the number went up. |
| 20% | Classroom contribution. How useful your questions and critiques are to other teams, in the round and in deep dives. Sitting silently through every other team's round, week after week, fails this component. |
| 20% | Demo Day. Clear explanation of the problem, product, and evidence; sound interpretation of results; and thoughtful answers to questions. Show what you learned within the three-minute pitch. |
| 20% | Final submission. Quality of the submitted product and supporting evidence; reasoning behind decisions; and reflection on limitations and lessons learned. See the submission requirements. |
| 0% | Traction. Deliberately zero. See below. |
Zero percent of your grade is traction. A team that runs five honest experiments and concludes their idea was wrong will outscore a team that quietly stalls and shows a polished demo at the end. This is deliberate. The moment your grade depends on your metrics, you will hide bad news, and hiding bad news is the one thing that makes this format useless.
Late weekly deliverables are not accepted. You get one drop; dropping a grade does not waive the required launch.
Prerequisites
There are no formal course prerequisites. Admission is by application and interview instead. You need to be able to build and ship a working product without being taught how, because this course does not teach it. Prior entrepreneurial experience is helpful but not required.
Policies
Using AI. Expected. You are building agents. In the final submission, note which parts were generated and which were written.
Intellectual property. Ownership depends on Columbia's intellectual property policies and any applicable employment, research, or funding agreements. Consult Columbia Technology Ventures about ownership and commercialization questions. If you plan to form a company, discuss ownership and equity with your teammates early. Incorporating a company or settling equity is not a course requirement.
Conflicts of interest. The instructor is a startup founder and advises and invests in early-stage companies. Visiting mentors and Demo Day judges are active investors. No mentor or judge will discuss investment or recruiting with any team until final grades are submitted. If you think a mentor has a conflict with your idea, tell the instructor and that mentor is recused from your deep dives and office hours.
Privacy. Applications and interview scores are education records, read by the instructor and the teaching staff. In course posts and presentations, anonymize user quotes and remove identifying customer details, confidential data, and credentials from screenshots and logs. If supporting evidence cannot be shared with the class, contact the instructor privately on Ed to arrange how to submit it.
Attendance. Each student may miss two class sessions. Notify the instructor privately on Ed in advance when possible. If you need additional absences or will miss a scheduled deep dive, mandatory 1:1, or Demo Day, contact the instructor to arrange how to make up the missed participation or assessed work. The point of being here is watching the other teams, not just presenting your own work.
Team conflict. Tell the instructor in week 4, not week 13. Almost every team problem is fixable early and almost none are fixable late.
Collaboration and Academic Honesty
Please read the Computer Science Department’s Academic Honesty Policies & Procedures and Columbia College and Columbia Engineering’s Academic Integrity policies.
What your team turns in must be your team's own work. Cite all sources you reference, including conversations with other teams and any tools or code you did not write. Using AI is expected and does not need citing beyond the note required in the final submission. Misrepresenting your metrics, your users, or what you shipped is the one form of dishonesty this course cannot tolerate, because the entire format depends on honest weekly reporting.