Harvard Fall Course Covers AI Agents and System Optimization

Summary

Harvard’s CS2680 Modern AI Systems: Agents and System Optimizations is a Fall 2026 course on building and operating agentic systems. It follows an agent from the interaction loop through the serving stack: students use Claude Code, design an agent with provided model API access, measure and optimize it, and then serve an open-weight model on course-provided GPUs. The systems portion covers batching, scheduling, routing, KV-cache and prefix reuse, quantization, speculative decoding, GPU and kernel work, and load balancing. The course has two parts, five assignments, a final project, and class participation, with no exam. Assignments progress from observing an agent in Mads-Lens and building its loop in Mads-Loop to token optimization, owning a serving engine, and optimizing the full stack. Prerequisites are CS61, at least one graduate-level computer systems course, and comfort with Python and PyTorch. The course meets Mondays and Wednesdays at Harvard’s SEC LL2.229; capacity was raised to 80 students, while auditing is allowed only when school policy and classroom space permit and does not include computing resources. Coursework uses Claude Code, model APIs, the course HPC, and optional AWS access, with support for OpenAI, Anthropic, and Gemini APIs.