CS2680 Modern AI Systems：智能体与系统优化课程公布 Fall 2026 安排
CS2680 Modern AI Systems: Agents and System Optimizations
哈佛 CS2680《Modern AI Systems: Agents and System Optimizations》Fall 2026 课程公布安排，容量增至 80 人，课堂时间为周一、周三 11:15–12:30，地点 SEC LL2.229。
CS2680 Modern AI Systems: Agents and System Optimizations
Fall 2026
Announcement
- Oct 8, 2026Assignment Leaderboard is live.
- Sept 4, 2026We have increased capacity to 80 students.
- Aug 28, 2026Audit policy: everyone is free to audit as long as it is permitted by school policy and the classroom has space. However, we do not have computing resources for auditing students.
- Aug 17, 2026Welcome to CS2680 Modern AI Systems: Agents and System Optimizations! Class meets Monday and Wednesday in SEC LL2.229.
Overview
This course follows one agentic system from the agent loop down through the serving stack. You first use a modern coding agent (Claude Code), then design your own agent using provided API access, measure it, and optimize it. Then you serve an open-weight model yourself and take responsibility for the layers underneath: batching and scheduling, KV-cache and prefix reuse, routing, quantization, and speculative decoding.
Class time and location
Monday and Wednesday 11:15am – 12:30pm SEC LL2.229
Prerequisites
CS61, and at least one graduate-level course in computer systems. Comfort with Python and PyTorch is assumed. See the background we assume.
Format. The course runs in two parts.
- Part I — Introduction to LLMs and agents. You will learn about how to use coding agents (Claude Code) and how to design an agent. You will also explore various agent architectures, use cases and optimizations.
- Part II — Systems for LLMs and agents. We will talk about LLM serving (batching, scheduling, routing, caching, quantization, and speculative decoding) and agentic systems.
The assignments walk down that stack in five steps. First, Mads-Lens1 lets you watch the agent work (Assignment 1). Second, you build the loop in Mads-Loop (Assignment 2). Third, you make every token count in Mads-Opt (Assignment 3). Fourth, you own the serving engine in Mads-Serve and run a model on the provided GPUs (Assignment 4). Fifth, you optimize the full stack in Mads-Stack (Assignment 5).
1 “Mads” abbreviates Modern Agent Design and Systems, the arc of this course, and nods to MaDSys, Harvard’s Measurement and Design of Systems research group, which teaches and supports it.
What this course expects. This course will not teach you every piece of knowledge you need for the assignments. It focuses instead on the skills to learn with, and points you in directions worth going. Expect to spend a significant amount of time learning and experimenting on your own after class. If you want an A, plan on it.
Grading.
- Assignment 1 Mads-Lens: watch the agent work10%
- Assignment 2 Mads-Loop: build the loop10%
- Assignment 3 Mads-Opt: make every token count20%
- Assignment 4 Mads-Serve: own the serving engine20%
- Assignment 5 Mads-Stack: optimize the full stack10%
- Project20%
- Class participation10%
- Bonus optional presentations up to 4% each; course feedback up to 6%; problems current AI cannot solve, 2% eachup to 10%
See policy page for details.
Staff
Instructor
TF
Advaith Ravishankar
TF
Charles Wang
TF
TF
Contact: cs2680-staff@g.harvard.edu Please use this address for all course-related questions. We only respond to emails from Harvard students. Due to the large number of emails we receive, we do not respond to external inquiries.
Office hours:
- Instructor: Appointment only
- TFs: See Ed + by appointment
Schedule
Tentative and subject to change. Rows with Students as the speaker are student-led: either a paper discussion (see the paper discussion page) or an assignment sharing session. Paper presentations are optional and can earn up to four bonus percentage points each.
| Week | Date | Speaker | Topic | Readings | Notes |
|---|---|---|---|---|---|
| Part I — Introduction to LLMs and agents | |||||
| 1 | 2026-09-02 | Juncheng Yang | Course overview and logistics [slide] | Assignment 1 out (Mads-Lens: Watch the agent work); Sign up for AWS | |
| 2 | 2026-09-07 | No class — Labor Day | |||
| 2 | 2026-09-09 | Juncheng Yang | Agents use I: How Claude Code Works [slide] | How I use LLMs; Best practices for Claude Code; Agentic Engineering Patterns; more reading → | |
| 3 | 2026-09-14 | Juncheng Yang | Agents use II: 100x Engineer with Multi-agents [slide] | Claude Code agent teams; Software is changing (again); Research acceleration: The view inside OpenAI; more reading → | |
| 3 | 2026-09-16 | Juncheng Yang | Agents design I: context management and tool design [slide] | ReAct; A practical guide to building agents; LLM Powered Autonomous Agents; more reading → | Assignment 1 due Sun Sep 20, 11:59pm |
| 4 | 2026-09-21 | Students | Student sharing I | What you learned from using an agent (Assignment 1): surprises, failures, lessons, insights, new tools/skills/plugins | Assignment 2 out (Mads-Loop: Build the loop) |
| 4 | 2026-09-23 | Juncheng Yang | Agents design II: multi-agent architecture, coordination, and communication [slide] | multi-agent research system; scaling agent systems; don’t build multi-agents; more reading → | |
| 5 | 2026-09-28 | Juncheng Yang | Agents design III: agent memory and reliability [slide] | where agents fail; MemGPT; more reading → | |
| 5 | 2026-09-30 | Juncheng Yang | Agent design IV: agent evaluation [slide] | Agent evaluation and benchmarks | Assignment 2 due Sun Oct 4, 11:59pm |
| 6 | 2026-10-05 | Harsh Nilesh Pathak (Snowflake) | Guest lecture: Developing and Evaluating Agents in Enterprise Setting | Assignment 3 out (Mads-Opt: Make every token count) | |
| 6 | 2026-10-07 | Juncheng Yang | Advanced agent design I: Cost and performance [slide] [Assignment3 slide] | Cost-effective agents | Pre-proposal due (optional and only submit if you want to discuss your ideas) |
| 7 | 2026-10-12 | No class — Indigenous Peoples’ Day | |||
| 7 | 2026-10-14 | Students | Advanced agent design II: agent safety, self-evolving harness | ||
| Part II — Systems for LLMs and agents | |||||
| 8 | 2026-10-19 | Juncheng Yang | Efficient LLM serving: basics I | Assignment 3 due Tue Oct 20, 11:59pm | |
| 8 | 2026-10-21 | Juncheng Yang | Efficient LLM serving: basics II | Assignment 4 out (Mads-Serve: Own the serving engine) | |
| 9 | 2026-10-26 | Juncheng Yang and Students | Assignment Feedback and Student sharing II | Common pitfalls in your submission; Assignment 3 lessons sharing. | |
| 9 | 2026-10-28 | Juncheng Yang | Efficient LLM serving: speculative decoding | Project proposal due | |
| 10 | 2026-11-02 | Juncheng Yang | Efficient agentic systems: scheduling and prefix caching | This lecture is earlier because of Assignment 4. | |
| 10 | 2026-11-04 | Juncheng Yang | Efficient LLM serving: pruning and quantization | ||
| 11 | 2026-11-09 | Xianzheng Kou (Meta) | Guest lecture: Why Infrastructure sets the pace of AI | Assignment 4 due Tue Nov 10, 11:59pm | |
| 11 | 2026-11-11 | Juncheng Yang | Efficient LLM serving: GPU and kernels | Assignment 5 out (Mads-Stack: Optimize the full stack) | |
| 12 | 2026-11-16 | Students | Efficient LLM serving: KV-cache optimization and model architectures | ||
| 12 | 2026-11-18 | Students | Student sharing III | Demonstrate your agent-engine co-design (Assignment 4) | |
| 13 | 2026-11-23 | Juncheng Yang | Efficient agentic system: routing and load balancing | Assignment 5 due Sun Nov 29, 11:59pm | |
| 13 | 2026-11-25 | No class — Thanksgiving recess | |||
| 14 | 2026-11-30 | Juncheng Yang | Efficient agentic systems: others | ||
| 14 | 2026-12-02 | Students | Student sharing IV | Showcase your full stack optimizations (Assignment 5) | |
| 14 | 2026-12-07 | Students | Final project poster and demo | Present your final project as a poster and a live demo. Location: SEC 1.402 from 9:00 AM to 1:00 PM |
Final project due Tue Dec 8, 11:59pm |
Reference pages
- Agent evaluation and benchmarks. Evaluation design, metrics, reliability, cost, and benchmark selection.
- Cost-effective agents. Techniques for allocating models, reasoning, context, collaboration, and serving resources across complete tasks.
Compute resources
Setup instructions for every environment the course uses are on the computing setup page. Compute for the assignments and the final project comes from:
- Claude Code. Required for Assignment 1.
- Model API access. Required for Assignments 2 and 3.
- The course HPC. Required for Assignments 4 and 5.
- AWS. Register in week one; using the account is optional.
来源：Hacker News · cs2680.seas.harvard.edu