A working demo can hide a brittle agent. The central lesson is to treat retrieval, control flow, data, and evaluation as one engineering problem: implement the machinery yourself first, then decide which abstractions are worth adopting.
What Sets It Apart
- The first assignment starts from a chat-completion call and a corporate email archive, without agent frameworks. Retrieval, tools, memory, and human approval remain visible design choices.
- The second assignment evaluates a supplied agent with code graders, an LLM judge, and error analysis, giving measurement its own engineering workload.
- The syllabus connects agent components to optimization, coordination, and safety, helping readers reason about interactions across the system.
- The project requires tasks, baselines, success criteria, and reproducible results. This makes evaluation a design input throughout development.
Who It Suits and the Tradeoffs
Great fit if you have an NLP background and want to justify agent architecture choices through implementation and experiments. Public readings, released slides, and assignment code offer a self-study path. Look elsewhere if you need an introductory programming course or unrestricted lecture videos: recordings require the course Canvas login, and materials are linked as released.
