MIT
AI Engineering. Fall 2026.
6.S978 Engineering AI Systems and Agents

Wednesdays 11:00am–1:00pm in 32-144


Schedule

We will add readings and lecture materials before each meeting.

DateTopic
From foundation models to AI systems
Sep 9Lecture 1Language Models as a Computational PrimitiveFor the first time in computing, it's possible to embed arbitrary "intelligent" functions inside software programs, by providing natural language specifications to pre-trained LLMs. We will study how LLMs are "grown" rather than engineered, how engineers should model their behavior, and why it thereby remains difficult to build reliable, efficient, and maintainable software systems around them.
Sep 16Lecture 2Requirements for Situated AI SystemsFoundation models supply broad capabilities, but what a specific AI system or agent should do depends on its downstream application and context. Engineers must elicit this context, some of which can only be specified in natural language. Other parts of the system's behavior must be expressed as code or learned gradually from feedback on specific decisions.
AI programming and agents
Sep 23Lecture 3Declarative AI ProgrammingA prompt to an LLM is too narrow a programming model for many AI systems. We will seek ways to express AI programs with declarative and typed contracts instead. Programs written this way can be automatically optimized from feedback or evaluations and stay more maintainable and portable as models change.
Sep 30Lecture 4Agent DesignWe will build on the previous lecture and dive more deeply into agent loops, a subset of AI programs in which the program gives the model consistent control but still decides what it sees and what it is allowed to do at each step. We will consider security and safety of AI agents, especially to threats such as prompt injection.
Evaluation and evidence
Oct 7Lecture 5Evaluating AI SystemsHaving built many systems and agents by this point, how do we know whether they work? We will start from qualitative coding and then discuss when a formal eval is required. We will also consider the range of useful evals, some of which consider average behavior like ML benchmarks whereas others represent something closer to software assertions. We will cover also the minimal indispensable statistics for useful evals.
Oct 14Lecture 6Modern Information RetrievalAI systems often need to search millions of documents. We will consider the difficult trade-offs in this problem and discuss the retrieval paradigms developed to tackle it (lexical search, cross-encoders, single-vector retrieval, and late interaction). We will then study retrieval problems that remain difficult even in 2026.
Oct 21Lecture 7Reasoning over Massive ContextsBeyond search, some tasks require our models to deeply reason about evidence spread across an entire collection of documents. We will study the trade-offs this presents and the different emerging paradigms around how agents navigate this problem.
Learning from evidence
Oct 28Lecture 8Downstream Learning for AI SystemsEngineers often improve an AI system by hand-crafting better prompts and programs. We will study learning algorithms that search over instructions, demonstrations, and the program structure itself from downstream rewards and even from qualitative feedback. We will consider how such algorithms respect the declarative programming model from earlier lectures.
Nov 4Lecture 9Downstream Learning in WeightsThere are cases where a situated AI system strikes its best trade-offs only by updating the model weights from downstream feedback. We will compare methods for supervised fine-tuning, reinforcement learning, and distillation and develop intuition for cases where the weights are the right parameters to learn vs. the program around the model.
Later topics and projects
Nov 11No classVeterans Day.
Nov 18Lecture 10Semantic Data SystemsTBD.
Nov 25Project StudioWe will test the claims, evaluations, and operating points behind the course projects before the final experiments.
Dec 2Lecture 11Continual LearningTBD.
Dec 9Project Presentations and SynthesisTBD.