Stanford changes the assignment: students must defend agent written code

Stanford's CS146S changed its assignments after coding agents could finish the previous year's homework with one prompt. Students now enter an unfamiliar open source project, propose a change and explain it to the person who could actually accept it.

Last year's homework gave way to an unfamiliar repository and a live maintainer

Latent Space describes the second run of The Modern Software Developer, taught by Mihail Eric. The 121 minute interview tours a ten week course covering the agent harness, context engineering, security, background agents and the software factory.

The most important change is the assignment. Open source maintainers bring real repositories into the course. A student must understand unfamiliar code, propose an intervention and defend the patch to the person responsible for merging it. The official course site lists 3 units and aims to teach students to design agent workflows, combine tools and skills and understand the limits of current coding agents.

A junior engineer's value moves from syntax toward earning trust

Once an agent can solve an isolated exercise, a traditional assignment no longer measures the student's work well. An unfamiliar repository adds what a prompt cannot easily bypass: incomplete context, local conventions, historical decisions and a person who can reject the change.

For companies, that resembles onboarding a junior engineer more closely. Producing a diff will be common. Identifying a worthwhile problem, verifying behavior and explaining a change so the team can trust it will matter more. The course tests the quality of oversight rather than resistance to AI.

One course demonstration does not establish a new industry standard

The source is an interview with the instructor, who also leads AI at Monaco. It offers detail about the course and his team's practice, but it is not a comparative study of student outcomes. It does not tell us whether graduates build more reliable software than students from a traditional course.

Open source review also measures only part of the work. A maintainer sees the patch and its explanation, but may not know how many decisions the student delegated or whether the student could diagnose a production failure alone. Assessment still needs independent diff reading, testing and incident work.

Accepted patches and independently found mistakes will settle the question

A strong signal will be changes that maintainers actually accept and the ability of students to revise them after critical review. More valuable still is whether a student can explain unexpected behavior, spot an error in an agent proposal and intervene safely when the agent loses context.

If such assignments spread, schools can stop racing tools with bans. Instead, they can place the student in front of a person who has the right to say that the patch does not belong in the project.

Lilith's verdict

An agent can write the student's patch. The student still has to face the maintainer, and that is where education begins beyond the submit button.