Stanford CS224V: Monica Lam Lectures on Agentic AI

CS 224V Agentic AI Fall 2026

Course Home · Schedule · Projects (Monica Lam, Fall 2026)

A Stanford Fall 2026 Agentic AI course built around a semester-long project, with lectures serving as the methodological skeleton: centered on reducing hallucination, accountable decision-making, hybrid knowledge reasoning, and long-horizon Agents, students either conduct foundational methods research or build a domain Agent; the final project accounts for 65% of the grade.

What does this course mainly cover, and how is it taught?

Instructor Monica Lam (Head TA Cyrus Zhou, etc.). Mon/Wed 15:00–16:20, classroom CoDa B80, attendance required; lectures are recorded on Canvas, but public slides are deliberately left blank to encourage in-person participation. 3–4 credits. Prerequisites: any one of LINGUIST 180/280, CS 124, CS 224N, CS 224S, 224U. Enrollment is capped to allow project supervision; currently waitlist only on Navigator (#2029). SCPD remote enrollment exists, but the university stresses it is not for scheduling conflicts; participation counts for 15%, with about 5% of local students attending class, plus weekend updates and weekly project mentor interactions (about 10% local and about 15% remote count toward this); TAs on Ed offer bonus. Assignments 20%, final project 65%. Projects are in pairs, with weekly follow-ups from on- and off-campus mentors; you may propose your own topic in the proposal document to recruit partners. Proposal form deadline is about 10/18 23:59 PT; auditors may email with subject "audit cs224v request".

The opening question is direct: LLMs hallucinate—how do we make useful, reliable Agents? The course grounds "computational thinking" in reading/writing retrieval algorithms, research knowledge curation and insight discovery, and accountable AI; the topic list includes: using RAG and formal task descriptions to suppress hallucination in QA/task settings; hybrid reasoning over databases, knowledge bases, and unstructured text; research knowledge organization and discovery; using formal methods to improve decision accuracy and interpretability; and automatic efficiency and accuracy improvements for long-horizon Agents.

Course learning arc following the schedule

9/23 Introduction: the hallucination problem, computational thinking, accountable AI, and the overall course design.

9/28 Knowledge curation: how to curate open domains with LLMs, iterating before writing, multi-perspective retrieval and reading, and interactive personalization. Systems: STORM / Co-STORM. Homework 1 released.

9/30 Research projects: research threads you can join and the AOS architecture, knowledge retrieval/curation/discovery; harnesses for scientific reasoning, accountable decision-making, dialogue management; domain Agents; harness training. Student intro form deadline.

10/5–14 Accountable AI and constrained retrieval: high-stakes decisions (e.g., clinical trial matching) require higher recall and precision than embedding similarity, and must provide faithful, reliable, complete justifications. Constraints are written as SMT (satisfiability modulo theories), using theorem provers as accountable decision-makers. Systems: SatIR, VERDICT. Follow-ups expand constrained retrieval with high recall and low false positives: normalization, logical entailment, the significance of missing-value imputation, and projecting from SMT to relational algebra, still using SatIR as the thread. Interspersed are student pitches and group proposal presentations; HW1 due about 10/6, HW2 and the formal proposal due about 10/18.

10/19 Low-hallucination RAG dialogue: how free-text-based dialogue can be grounded—retrieval, summarization, verifying generation, response generation—and discussion of factuality and evaluation methods in persuasive settings. System: WikiChat.

10/26 Querying knowledge bases and graphs: opening massive fact bases like Wikidata with natural language. Semantic parsing + ReAct. Systems: SPINACH (Wikidata), DataTalk (FEC, etc.).

10/28 Data-driven deep research Agents: extracting insights from tabular data and reproducing published findings; drawing on exploratory data analysis and data storytelling to rewrite deep research over structured data as a "thesis-driven" analysis process. System: DataSTORM.

11/2 Hybrid data querying: what to do when a question needs both structured and unstructured answers. SUQL is introduced as a query language combining information retrieval (natural language) and database queries (SQL), with discussion of language design and evaluation.

11/4 Long context, long-horizon deep research: reading, comparing, and analyzing across a large corpus of documents when context is never enough. The path is to automatically build schemas, build databases, and align data before answering. System: SLIDERS.

11/9 Low-hallucination task-oriented Agents: Genie Worksheet as a high-level task specification language, where the author mainly describes "the task to be completed" and the low-level dialogue implementation is left to the language runtime; covers formal dialogue states, semantic parsing, dialogue state tracking and response generation, and how to build reliable Agents at scale across domains.

11/11 Coding Agents: using examples like Claude Code and OpenAI Codex, decomposed into LLM + tools + harness, and discussing meta-optimization of the harness.

11/16 VLM transcription: the challenges of language, layout, and digitization quality in ancient documents/printed materials, and fine-tuning vision-language models to make historical documents searchable and readable.

11/18 Introduction to LLM training: where instruction-following models and training data come from.

After that is the Thanksgiving break, with group final presentations on 12/2. Past offerings from Fall 2025 / 2024 / 2023 are available for comparison.

What this course emphasizes

The whole schedule is almost "one system per lecture": it does not stop at prompt tricks but ties Agents to curation pipelines, SMT/constrained retrieval, verified generation, ReAct over KBs/graphs, thesis-driven analysis over tables, SUQL hybrid querying, automatic database building over long documents, the task specification language Genie, and coding harnesses. Projects carry heavy weight and mentors follow up weekly, making the format more like a supervised Workshop.