Stanford University's CS146S, a course on AI-native software engineering, is in its second run, and instructor Mihail Eric has changed its assignments in response to a practical problem: a coding agent can now complete last year's homework in a single prompt. Rather than set exercises that an agent can automate, the course has brought open-source maintainers into the curriculum. Under the revised format, students must learn an unfamiliar codebase and propose a change to it, a workflow closer to real open-source contribution than to standalone programming exercises. The shift moves the emphasis from producing code on demand toward understanding existing systems and deciding what should change. The discussion, published as part of the Latent Space One Million AI Engineer special, centres on how the course is designed and how it is run. For Australian businesses and professionals, the change is mainly one of training design rather than new cost, risk or regulation. It does not alter obligations under the Privacy Act. The practical lesson is that if your team hires or trains junior developers on the assumption that AI tools will handle routine coding, the skills worth building are reading unfamiliar code, scoping changes and reviewing AI-generated work. A low-cost test would be to have a new hire take a small, real change through review in an existing codebase, with AI assistance permitted, and see what that reveals about their judgement.