tl;dr: A CMU professor teaching Machine Learning in Production redesigned assessments after AI agents could complete all his assignments: written reflections and reading quizzes were replaced with mandatory 15-minute in-person TA check-ins, exams were weighted more heavily, and assignment scope was scaled up (e.g., extending a 500k-LOC Zulip codebase instead of a toy Flask app). Grading itself is now partially automated via LLM-as-a-judge with human review, freeing TAs for face-to-face interaction. He acknowledges these changes violate some evidence-based pedagogy (fewer low-stakes assessments) but sees no alternative as models keep improving.