CS2680 Modern AI Systems: Agents and System Optimizations
Computing setup

Use the smallest environment that does the work

You will use Claude Code throughout the course. Add AWS when you deploy Assignment 1, move to the HPC when a run needs more time or a GPU, and use CloudLab only when a final project needs control of an entire machine. These tools solve different problems; setting up all four at once buys you nothing.

Do the approval-dependent work in week one. Request the course Claude Code account, get HPC access, and register your AWS account with the course. You can install software later; you cannot make an account approval happen the night before a deadline.
Required · every assignment
Claude Code

Start here. Harvard FAS provides the account, so do not buy a subscription. Finish the request, install, login, and telemetry setup, then learn where Claude Code stores the session files you submit with every assignment.
GPUs and long-running agent sessions
HPC

Use an interactive Slurm job when a run is too long for your laptop, and use the course GPU partition for Assignments 4 and 5. The login node is for submitting jobs and moving files; do not run the agent there.
Required for Assignment 1
AWS

Deploy the agent console on a small EC2 instance; no GPU is needed. Register the account with the course, protect the endpoint, set a billing alarm before launch, and stop the instance when you are finished.
Optional · project work
CloudLab

Choose bare metal when a project must change the kernel, network, or storage stack, or when a measurement cannot tolerate a co-tenant. Otherwise, the HPC is the simpler and faster choice.

Assignment by assignment

Work Use Why
Assignment 1 Claude Code + AWS or CloudLab Build with the course agent account and deploy web app. No GPU required.
Assignments 2 Assignments 3 Laptop/HPC + Harvard OpenAI/Gemini/Anthropic API The model runs behind a commercial API. Any machine is enough; use AWS or HPC job when the experiment should run for hours.
Assignments 4 HPC GPU (RTX PRO 6000 Blackwell) You serve the model yourself. Use the course gpu-cs2680 partition.
Assignments 5 HPC and/or API + AWS/Cloudlab You serve the agent using the provided GPU, you will optimize your model and serving engine, but you can also use a small number of API calls.
Final project HPC by default; CloudLab when necessary Use the HPC for model and GPU work. Use CloudLab only when the research question depends on bare-metal control or isolation.

If a project needs hardware that neither environment provides, ask the instructor while the scope can still change. Compute can sometimes be arranged; last-minute capacity cannot be promised.

Five rules for every machine

  • Keep secrets out of agent sessions. A transcript can contain files the agent read and output its commands printed, not just your prompts. Follow the scan-before-you-archive checklist every time.
  • Configure telemetry on each machine. The course settings live in ~/.claude/settings.json. Your laptop, the HPC, and an EC2 instance each need the telemetry block.
  • Name the system behind every result. Report the GPU, driver, serving engine and version, model, and workload. Latency or throughput without that context cannot be reproduced or compared.
  • Release idle resources. Exit the Slurm job, stop the EC2 instance, and terminate the CloudLab experiment. Shared hardware and paid instances should never sit idle by accident.
  • Escalate access problems early. An account problem in week one is fixable. The same problem on the due date is still an account problem, not an extension plan.