Harvard's CS2680 teaches modern AI agents from Claude Code down to KV cache and quantization
CurieuxExplorer · x · 2026-10-12
Harvard's new course CS2680 Modern AI Systems: Agents and System Optimizations (Fall 2026) is built around one question: how does a modern AI agent run from the application layer all the way down to the hardware stack?
- Structure: Part I covers LLMs and agents — using Claude Code, designing your own agent, architectures and optimizations. Part II covers the serving stack — batching, scheduling, routing, KV-cache and prefix reuse, quantization, and speculative decoding.
- Five assignments walk down that stack: observe an agent at work (Mads-Lens1), build the agent loop (Mads-Loop), optimize token costs (Mads-Opt), then build your own harness and optimize the inference stack, including self-serving an open-weight model.
- Logistics: capacity expanded to 80 students; grading is five assignments plus a final project, no exam; free auditing allowed but auditors get no compute. Prerequisites: CS61 and a graduate-level systems course, plus Python/PyTorch fluency.
It's being shared as a practical curriculum for anyone building agents, harnesses, or AI infrastructure. Note: the attached image is an AI-generated illustration unrelated to the actual course, per the official site.
More from coding & agent
- Ice Cubes dev uses Codex to clear open-source repo backlog of bumps and fixes — Dimillian · 2026-10-12
- SEO team uses Grok bot groups as web crawler, sometimes beating ScreamingFrog locally — gaganghotra_ · 2026-10-12
- EXL3-QLoRA lets you fine-tune 2-8bpw quantized models, beating BNB NF4 on loss — Unstable_Llama · 2026-10-12
- Linus Torvalds is using Claude Code on a hefty Linux kernel performance improvement — PointmanW · 2026-10-12
- Claude-driven Blender 3D modeling from images makes this dev bullish on inspection gear — jwt0625 · 2026-10-12
- Work2Skill Turns Human GUI Demos into Reusable Agent Skills, Beating No-Skill Baselines — nex-agi · 2026-10-12