STML 2026 — LLM Agents
Special Topics in Machine Learning: LLM Agents · Graduate course · 15 weeks
Week by week
Select a week below to open its HTML chapter, lab, homework, and presentation papers. Every week’s topic and papers are listed from the start; the materials for a week are released in the week it is taught.
| Wk | Topic | Lecture | Lab |
|---|---|---|---|
| 1 | Course intro · What is an agent? | agent = control flow decided by model output; components; autonomy | first API calls · prompting fundamentals |
| 2 | Prompting & reasoning | chain-of-thought · self-consistency · test-time compute | CoT & exemplars, measured on a math evalset |
| 3 | Tool use | function calling · tool schemas · the capability boundary · code execution | functions into tools (Ng M3) |
| 4 | The agent loop (ReAct) | reasoning–action interaction · worked prompt · controlled comparison · loop execution | the loop built by hand |
| 5 | Reflection & evaluation | review and revision · evalsets · code/LLM judging · error analysis | draft–review–revise · general evaluation (Ng M2+M4) |
| 6 | Multi-agent systems | specialization · role inputs and outputs · communication patterns | multi-agent pipeline (Ng M5) |
| 7 | Planning & search | decomposition · plan–execute–replan · ReWOO · ToT | research-agent workflow (Ng final project) |
| 8 | Midterm exam | written, W1–7 | — |
| 9 | Retrieval augmentation (RAG) | embeddings · pipeline · agentic RAG | retriever for the final project |
| 10 | Context engineering · Memory | Lost in the Middle · select/order/compress · MemGPT · Mem0 | memory · graph integration |
| 11 | Reasoning models & RL | o1/R1 · STaR · RLVR · GRPO | thinking-budget measurement |
| 12 | Inference economics · Benchmarks | cost–accuracy trade · routing · Pareto · pass^k | routing + cost logging · eval harness |
| 13 | Safety & security · Retrospective | indirect prompt injection · defense layers | guardrails · red team |
| 14 | Final presentations | team demos + eval scorecard defense | — |
| 15 | Final exam | written, whole semester | — |
Ordering principle: concept dependency — every week is understandable using only what came before it. Materials appear on each week’s page as the semester progresses.
The course in one sentence
You will learn how LLM agents work by building one, week by week, and by reading the papers that invented each part. The final deliverable is a research-assistant agent that searches and synthesizes answers over this course’s own paper corpus.
Weekly rhythm (3 hours)
| Block | Time |
|---|---|
| Theory lecture | 30–40 min |
| Paper presentations ×2 | 50 min |
| Break | 10 min |
| Lab (Colab notebook) | 80 min |
| Wrap-up · next-week preview | 10 min |
Labs run in Google Colab with your own API key — no local installation. Week 4 uses the OpenAI SDK to make the agent loop visible; other early labs use aisuite, and the later labs use LangChain and LangGraph. Use the Open in Colab links where provided, or download the .ipynb and use File → Upload notebook in Colab. The early agent labs adapt Andrew Ng’s Agentic AI (DeepLearning.AI), and the later sequence follows LangChain Academy and official RAG tutorials. Follow each week’s instructions for in-class work and homework.
Before the first class
- Read the API Setup Guide and issue your own API key. The Week 1 lab needs it in class.
- Bring a laptop that can open Google Colab in a browser.