Stanford CS146S shifts AI coding assignments to open-source contributions

Mihail Eric teaches CS146S, Stanford's Modern Software Developer course, and leads AI work on Monaco, a role combination that let him observe both how students learn AI coding and how engineering teams use coding agents in practice. When CS146S first ran, it was built around nine assignments and a capstone project. One early exercise had students run an AI code review tool on a repository and evaluate the feedback it produced. In the following term, Eric concluded that many of those exercises had collapsed into single-prompt tasks that an agent could now complete in one shot. He also polled a group of Stanford sophomores and juniors about how they code, and found that all of them already used agents in their workflow. The open question became what assignments should actually measure when writing code on its own no longer demonstrated learning. The next iteration restructured part of the course around real open source contributions: students submit patches to open source projects, and the projects' maintainers review them. Compared with closed classroom exercises, this setup adds requirements that cannot be satisfied by code generation alone. Students must confirm that maintainers actually want the proposed feature, understand why existing code behaves the way it does, assess what other behaviors a patch might affect, and then explain, revise, and refine their proposals in response to maintainer feedback. As the course puts it, an agent can build a feature in an afternoon, but deciding whether it belongs in the project still requires communication and judgment. The course also introduced exercises such as explain-diff, in which students explain code changes rather than only generating them. No information appears yet on how the redesigned assignments have performed in practice or on what comes next.