Berkeley Large Language Model Agents Course

What the course is

CS294/194-196, Large Language Model Agents, ran at UC Berkeley in Fall 2024, taught by Professor Dawn Song with Xinyun Chen of Google DeepMind. The public MOOC version made every lecture recording, slide deck and quiz available online for free, and remains on the course site. A follow-up Spring 2025 course on more advanced agent topics was run by the same group.

The lecture series

Most lectures are given by guest researchers from industry and academia, which is the course's main attraction. Topics and speakers include:

• LLM reasoning, Denny Zhou (Google DeepMind)
• History and overview of LLM agents, Shunyu Yao (OpenAI)
• Agentic frameworks and multimodal systems, Chi Wang and Jerry Liu
• Enterprise AI trends and retrieval-augmented generation, Burak Gokturk (Google)
• Compound AI systems and DSPy, Omar Khattab (Databricks)
• Software development agents, Graham Neubig (CMU)
• Enterprise workflow agents, Nicolas Chapados (ServiceNow)
• Neural and symbolic decision-making, Yuandong Tian (Meta AI)
• Robotics and foundation models for humanoids, Jim Fan (NVIDIA)
• Open-source foundations, Percy Liang (Stanford)
• Agent safety and responsible scaling, Ben Mann (Anthropic) and Dawn Song

Before you start

This is a graduate-level seminar, not a coding tutorial. You will get more from it if you already:

• Understand how language models are used: prompting, APIs, embeddings and retrieval at a basic level.
• Program comfortably in Python.
• Can read a research paper, or are willing to learn, since lectures refer to papers constantly.

If you have never called an LLM API, build a small chatbot and a simple retrieval app first.

How to get the most from it

• Watch in order for the first three lectures, then pick topics that match your interests.
• Read the key paper behind each lecture; lectures summarise, papers teach.
• Build something per lecture: a reasoning prompt technique, a tool-calling agent, a small DSPy program, a coding agent on a toy repository.
• Take notes on open problems speakers mention. They make good project ideas and conversation topics with researchers and hiring managers.

At one lecture a week plus reading and building, expect about three months.

Certificates and the hackathon

During the live run, MOOC learners could earn certificate tiers (Trailblazer, Mastery, Ninja, Legendary and Honorary) by completing quizzes, a written article, lab assignments or a hackathon project. That round has finished, so treat the course now as free learning material rather than a credential.

Where it stops

It reflects the field as of late 2024. Agent tooling has moved quickly since, so pair the foundations here with current framework documentation. It also teaches ideas more than production engineering such as evaluation pipelines, cost control and deployment.

Common questions

Is it free? Yes, all materials are free online.

Can I still get a certificate? Certificates were tied to the Fall 2024 run; check the site for any current offering.

Is it too advanced for an undergraduate? For a strong final-year student with some LLM experience, it is manageable. Otherwise, start with a practical course first.

Best for: engineers and advanced students who want the research ideas behind modern AI agents from the people who developed them.