STML 2026 · All chapters
Read the chapters in order, or choose a topic below.
Chapter 1 · What is an Agent?
Follow a library-room reservation through a failed attempt and a confirmed alternative. Define the system's components and compare a predefined workflow with model-directed decisions.
Chapter 2 · Prompting & Reasoning
Use one arithmetic task to compare instructions, worked examples, and sampled solutions. Explain why a valid-looking solution or a winning vote can still be wrong.
Chapter 3 · Tool Use
Define a clock function and describe its purpose and parameters to the model. Follow the request, execution, and returned result, then connect two tools and introduce Toolformer.
Chapter 4 · The Agent Loop and ReAct
Follow a question through two searches. Connect an observation to the next action, then apply the same loop to SQL.
Chapter 5 · Reflection & Evaluation
Understand how critique and revision improve existing work, and how explicit criteria measure quality. Study reflection with a chart example and evaluation with invoice extraction and summary coverage.
Chapter 6 · Multi-Agent Systems
Learn what multi-agent systems are and why they are used. Study collaboration and communication structures, then follow a travel plan through research, review, and revision.
Chapter 7 · Planning & Search
Build a plan from a goal, connect tasks through their required results, and revise the work when execution reveals a problem. Study ReWOO and Tree of Thoughts, then inspect a research agent system.
Chapter 8 · Retrieval-Augmented Generation
Use a four-document toy corpus to follow a question through chunking, embedding, and top-k retrieval to a grounded answer. Separate retrieval errors from grounding errors, then connect the next search to agentic RAG and Self-RAG.
Chapter 9 · Context Engineering
Follow a long conversation with the reading-list assistant whose history outgrows its budget. Explain the capacity and use limits of context, then select, structure, and compress the input and compare context policies by input, answer, and token cost.