CS 224G: Building and Scaling LLM Applications
by Stanford University
Stanford's Winter 2026 product-engineering course on the LLM application stack — every lecture deck is a public PDF.
Overview
CS 224G ran in Winter Quarter 2026 from 6 January to 12 March, taught by John Whaley and Jan Jannink, both CS PhDs and serial founders with exits. The format is ten weeks structured as four two-week sprints, teams of two to five, graded on bi-weekly demos (45%), a Demo Day presentation (45%) and participation (10%); Demo Day was held 19 March 2026 at Stanford's CoDa building with roughly nineteen projects shown to VCs, faculty and press. The lecture arc: the 2026 reasoning-model landscape (o1/o3, DeepSeek R1, Claude) with chain-of-thought and cost/latency tradeoffs; two consecutive lectures on context engineering and RAG covering the context stack, memory systems, prompt-injection defence, instruction strata, call parameters and agentic loops; agentic workflow design patterns (ReAct, multi-agent, orchestration frameworks) with a linked Colab notebook; code generation and coding agents benchmarked against HumanEval and SWE-bench; agentic orchestration with Pydantic AI plus a public base-agent repo; data strategy and the memory layer as competitive moat; a realtime voice AI workshop over WebRTC with its own repo; and AI safety, guardrails, Constitutional AI and the 2026 regulatory landscape. Guest sessions cover AI-native company building, applying the Theory of Constraints to non-deterministic systems, pitching and how investors qualify AI startups. Be clear on the limits: the lecture PDFs under /lectures/ are freely downloadable, but the course is deliberately not recorded, and the sprints, Slack and Canvas are for enrolled students only. Winter 2024 and Winter 2025 editions are archived on the same site.
At a Glance
- Topic
- Agentic
- Level
- Intermediate
- Format
- Course
- Cost
- Free
- Duration
- 10 weeks, 20 lectures — ~12-18 hours to read all publicly posted Winter 2026 slide decks
- Provider
- Stanford University
- Hands-on
- Yes — code/exercises
- Certificate
- None
What You’ll Learn
- ✓Assemble a production context stack from instruction strata, retrieval, memory and call parameters
- ✓Design agentic workflows with ReAct and multi-agent patterns and pick an orchestration framework
- ✓Defend an LLM application against prompt injection at the context layer rather than the prompt
- ✓Reason about reasoning-model cost and latency tradeoffs when choosing between o-series, R1 and Claude
- ✓Build production-ready agents with Pydantic AI starting from the course's public base-agent repository
- ✓Ship a realtime voice application over WebRTC using a native realtime pipeline
- ✓Diagnose non-deterministic systems with the Theory of Constraints to find the binding failure first
- ✓Apply guardrails and Constitutional-AI techniques against the 2026 regulatory landscape
Highlights
- •Every Winter 2026 lecture deck is a downloadable PDF on the public course site — no enrolment, login or paywall
- •Taught by two startup founders with exits, so the framing is shipping products rather than publishing papers
- •Guest lectures come from practitioners at working AI companies, each contributing their own deck, notebook or repo
- •Two full lectures plus a Colab notebook go deep on context engineering, a topic most application courses cover in one slide
- •Winter 2024 and Winter 2025 editions are archived alongside 2026, so you can see how the stack advice has shifted year over year
Who It’s For
Best For
- ✓Engineers moving an LLM prototype toward something with real users and real cost
- ✓Founders and technical leads scoping an AI product in a ten-week sprint format
- ✓Self-learners who want a current, opinionated survey of the 2026 LLM application stack
Prerequisites
- •Python proficiency at roughly the level of Stanford CS 106B
- •Machine learning fundamentals equivalent to Stanford CS 221 or CS 229
- •Hands-on experience with at least one of PyTorch, TensorFlow or JAX
FAQ
What is CS 224G: Building and Scaling LLM Applications?
CS 224G is Stanford's project-based course on building and scaling LLM applications, and its Winter 2026 slide decks are publicly hosted on the course site. It targets engineers who can already write Python and train a model and now want the production layer: context engineering, agent orchestration, evaluation of non-deterministic systems, realtime voice, guardrails and cost/latency tradeoffs. Reading the twenty decks gives you a current, opinionated map of the 2026 LLM application stack from two people who ship products.
Is CS 224G: Building and Scaling LLM Applications free?
CS 224G: Building and Scaling LLM Applications is free to access.
What level is CS 224G: Building and Scaling LLM Applications for?
CS 224G: Building and Scaling LLM Applications is aimed at a intermediate audience. Recommended background: Python proficiency at roughly the level of Stanford CS 106B, Machine learning fundamentals equivalent to Stanford CS 221 or CS 229, Hands-on experience with at least one of PyTorch, TensorFlow or JAX.
How long does CS 224G: Building and Scaling LLM Applications take?
Expect roughly 10 weeks, 20 lectures — ~12-18 hours to read all publicly posted Winter 2026 slide decks. Most learners work through it at their own pace.
What will I learn from CS 224G: Building and Scaling LLM Applications?
You'll learn: Assemble a production context stack from instruction strata, retrieval, memory and call parameters; Design agentic workflows with ReAct and multi-agent patterns and pick an orchestration framework; Defend an LLM application against prompt injection at the context layer rather than the prompt; Reason about reasoning-model cost and latency tradeoffs when choosing between o-series, R1 and Claude; Build production-ready agents with Pydantic AI starting from the course's public base-agent repository; Ship a realtime voice application over WebRTC using a native realtime pipeline; Diagnose non-deterministic systems with the Theory of Constraints to find the binding failure first; Apply guardrails and Constitutional-AI techniques against the 2026 regulatory landscape.
Topics
Sources
This page was written from 4 sources, 1 on domains other than web.stanford.edu.