Harvard CS2680 course traces agent loops down to inference engines

Meng Shao reports on Harvard's fall 2026 CS2680 course, taught by Juncheng Yang, which follows one agentic system from the agent loop down through the serving stack.

According to Shao, the course has five assignments that progress from observing an agent, to writing a minimal harness, to optimizing around a black-box API, to self-hosting an inference engine, and finally full-stack optimization with measured results.