Modern AI Systems: Agents and System Optimizations (CS2680)
Juncheng Yang · Harvard SEAS · Fall 2026freeSource ↗Assignments — all five, with dates and weights ↗Starter repo (HarvardMadSys/cs2680_assignments) ↗Assignment 1 · Mads-Lens spec ↗Assignment 2 · Mads-Loop spec ↗Assignment 3 · Mads-Opt spec ↗Computing setup — what each assignment needs ↗Course blog ↗
Added 2026-10-12 at Connor’s direction, off the ladder on the same terms as stanford-cs146s. It follows one agent system from the loop down through the serving stack: use Claude Code, build your own harness, cut its token bill, then serve an open-weight model yourself and optimise the whole stack. That is the token-to-task path walked in a single course, so it sits across harnesses, evals and serving rather than on one rung. Ungated because its prerequisites (CS61, a graduate systems course, Python and PyTorch) are not units here. Scope counts the sixteen instructor lectures; the five assignments are the actual work and are listed as segments.
Third in the lockstep, sharing the technical slot with stanford-cs146s and agentic-ai (Connor, 2026-10-12). Each round: CS146S lesson N → 11-768 lecture N → CS2680 lecture N. Assignments go with their week.
Environment
Setup is a real gatepartial~/dev/study/harvard-cs2680/verified
Starter repo cloned (A2 and A3 published; A1 has no starter; A4–A5 unreleased). The eight public slide decks are under slides/, and the A1–A3 spec pages under specs/. A1 runs today on Connor’s own Claude Code at zero cost. A2 and A3 need Docker (not installed) and a qwen3.6-35b-a3b endpoint: the course proxy is students-only, so the choices are a local quantised model on this 36 GB Mac or a paid API, and paid needs Connor’s per-spend approval. README-local.md has the table.
Segments
The named subset — the scope above is the count| Ref | Title | Done | Artifacts | Note |
|---|---|---|---|---|
| A1 | Assignment 1 · Mads-Lens — a web page that drives Claude Code headless and streams the trajectory live: tool calls, subagent tree, failures, continuation, cost | · | NBBLDGMEBTL0/0 | Course due 9/20 (a pace, not a deadline). Overlaps CS146S A1: both read the same agent from outside. |
| A2 | Assignment 2 · Mads-Loop — your own coding-agent harness on ten SWE-bench Pro tasks, with tool schemas, context management, retries, termination and graceful failure | · | NBBLDGMEBTL0/0 | NEEDS — Docker plus a qwen3.6-35b-a3b endpoint (local or paid). |
| A3 | Assignment 3 · Mads-Opt — the harness competition: hold task success fixed, drive cost and latency down | · | NBBLDGMEBTL0/0 | NEEDS — Same as A2. The leaderboard is students-only. |
| A4 | Assignment 4 · Mads-Serve — serve a 7–8B open-weight model yourself, measure the quality loss, recover it through the harness | · | NBBLDGMEBTL0/0 | NEEDS — A GPU. Spec not yet released. |
| A5 | Assignment 5 · Mads-Stack — profile the full path, state a hypothesis, optimise across layers, re-run the same workload | · | NBBLDGMEBTL0/0 | NEEDS — A GPU. Spec not yet released. |
Artifacts
Consume → do → output- The assignments, done — a trajectory viewer, a harness, an optimised harness, then a self-served model and a full-stack pass as they are releasedNot yet — nothing has landed for this slot.
- What it cost to hold task success fixed while cutting tokens — the A3 write-up, measuredNot yet — nothing has landed for this slot.
Feeds
What closing this sharpensNo notes on this unit yet.
A note lands here when a Learning Log entry tagged study-harvard-cs2680 is published.
Connor is not enrolled, and the course gives auditors no compute: no Claude Code licence, no course API key, no HPC GPUs, no leaderboard. A3’s leaderboard (part of its grade) is students-only, and A4–A5 assume the course’s RTX PRO 6000 Blackwell partition, which has no stand-in here yet. A3 and A5 are submitted as published blog posts, which fits /notebook directly.