Stanford CS329Z Introduction — What Are Agentic Systems?
Stanford Online
Course logistics 0:05
CS 329Z is introduced as Engineering AI Agents. The instructors say the class will use an ad discussion list for contact, with the website holding lectures, slides, and homework. They ask you to use AI to learn, not during lecture, and recommend some NLP or language model background. Students get compute credits, and the course warns that AI can lift homework scores while hurting exam learning.
From models to agents 8:02
The lecture traces agents from old ideas of software that acts on its own to the modern loop of perceiving, acting, and learning from rewards. It then shows what changed with large models: few-shot prompting, chain-of-thought reasoning, web browsing, and tool use. A short timeline covers ReAct, ChatGPT, GPT-4, BabyAGI, and AutoGPT, then notes that newer models and benchmarks have pushed agent performance and efficiency higher.
Model cost and progress 19:31
Longer thinking usually helps. The speaker says API cost tracks with how many tokens the model digests, so better results often come with higher cost. Why newer models beat earlier ones is still unclear. Possible reasons include better data, better distillation, or other training changes, and the details will be taken up later.
Agent architecture 22:01
An agent built on a language model needs planning, memory, tools, actions, and environment feedback. The model thinks through a task, stores what has happened, and acts with tool calls such as search or movement. Reasoning helps because it can explain an action, make the same step work in new cases, and reduce hallucinated answers in loops like React.
Reflection and debate 38:02
Self-reflection is described as compressing a trajectory of events into something shorter. That is different from classic distillation, which trains on outputs from other models. The discussion then turns to multi-agent debate. Several agents solve the same task, read one another’s reasoning, and argue over it. That can help when models are weaker, though the main value today may be in efficiency, access control, and dividing sensitive work across agents.
Memory and tools 43:02
Memory is needed because long tasks do not fit in a context window, and because a model must still pick the right detail from a large pile of information. The transcript separates episodic memory, semantic memory, and procedural memory. It also outlines tool use, from early systems that learned when to call an API to newer work like Gorilla, which retrieves from many APIs and lowers hallucinations by handing hard tasks to outside tools.
Workflows and agents 56:00
A key design choice is who decides the steps and the memory. A workflow fixes the order. The model works inside each stage, but it does not choose the next one. An agent is more flexible. It decides which files to open, what to run, and when it is done. Workflow systems can still be strong, as AlphaCode showed with millions of candidate programs and a high rank among human competitors. Retrieval augmented generation, or RAG, is another workflow pattern. It connects a model to outside sources, such as a weather API, so answers stay more factual and consistent.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.

