Agentic AI MOOC (UC Berkeley, Fall 2025)
by UC Berkeley RDI
Twelve two-hour lectures on agentic AI from the people actually shipping it — free, recorded, and tiered by how much work you want to do.
Overview
Berkeley RDI's Fall 2025 Agentic AI MOOC is the third instalment of Dawn Song's LLM-agents course sequence, following the Fall 2024 LLM Agents MOOC and the Spring 2025 Advanced LLM Agents MOOC, and it assumes you have absorbed the fundamentals those covered. The format is twelve two-hour Monday lectures recorded at Berkeley between 15 September and 8 December 2025, each delivered by a working researcher rather than course staff: Yann Dubois (OpenAI) on LLM agents overall, Yangqing Jia (NVIDIA) on how system design has evolved from an AI engineer's perspective, Jiantao Jiao (NVIDIA) on post-training verifiable agents, Weizhu Chen (Microsoft) on the practical difficulties of training agentic models, Noam Brown (OpenAI) and Oriol Vinyals (Google DeepMind) on multi-agent systems, Sida Wang (Meta) on predictable noise in LLMs, James Zou (Stanford) on agents that automate scientific discovery, Clay Bavor (Sierra) on deploying real-world agents, Peter Stone (UT Austin / Sony AI) on embodied autonomous agents, and Dawn Song herself on agentic AI safety and security. A course-staff lecture covers agent evaluation and the project brief. Assessment is tiered rather than pass/fail: a Trailblazer certificate for lecture quizzes plus a Learning and Sharing assignment, Mastery for submitting an AgentX-AgentBeats competition project, Legendary for placing as a winner or finalist in that competition, and Honorary for exceptional Discord community contribution. All coursework was due 31 January 2026; the signup page reports a community of more than 32,000, and the lecture videos remain freely available in the syllabus.
At a Glance
- Topic
- Agentic
- Level
- Advanced
- Format
- Course
- Cost
- Free
- Duration
- 12 lectures x ~2h, Sep 15 - Dec 8 2025; free to audit on video, self-paced
- Provider
- UC Berkeley RDI
- Hands-on
- Yes — code/exercises
- Certificate
- Available
What You’ll Learn
- ✓Trace how agent system design evolved from single-call prompting to verifiable post-trained models
- ✓Evaluate agents properly, including what lab benchmarks systematically fail to measure in production
- ✓Compare multi-agent coordination approaches from two labs that built them at frontier scale
- ✓Understand post-training and reinforcement learning recipes used to make agentic models reliable
- ✓Apply agents to scientific discovery workflows including hypothesis generation and experiment design
- ✓Identify the latency, robustness, safety and user-trust failures that only surface after deployment
- ✓Reason about embodied and interactive agents where the environment, not text, is the interface
- ✓Assess agentic AI security threats and the defences currently available against them
Highlights
- •Every lecture is given by a practitioner from OpenAI, Google DeepMind, Microsoft, Meta, NVIDIA, Stanford or Sierra — not by teaching staff summarising their papers
- •Four certificate tiers let you take it as a lecture series or as a full competition project, without penalising the lighter path
- •The AgentX-AgentBeats competition gives the course a real artifact to build rather than a graded quiz
- •Third instalment of a sequence, so it skips the introductory material most agent courses spend half their runtime on
- •Completely free with no paywalled tier; all lecture recordings stay published in the syllabus after the term ended
Who It’s For
Best For
- ✓Engineers who already ship LLM agents and want the research context behind the failure modes they hit
- ✓ML researchers moving from model training into agent post-training and evaluation
- ✓Technical leads choosing between single-agent and multi-agent architectures for a production system
- ✓Anyone who finished the Fall 2024 or Spring 2025 Berkeley LLM Agents MOOCs
Prerequisites
- •Working knowledge of large language models, prompting and tool calling
- •Comfort reading ML research papers — lectures cite them without recapping
- •Python and prior experience building at least one LLM application, for the project tiers
FAQ
What is Agentic AI MOOC (UC Berkeley, Fall 2025)?
Dawn Song's Fall 2025 MOOC out of UC Berkeley's Center for Responsible Decentralized Intelligence, the third in her LLM-agents course sequence. Twelve guest lectures from OpenAI, Google DeepMind, Microsoft, Meta, NVIDIA, Stanford and Sierra cover agent architectures, post-training, evaluation, multi-agent systems, embodied agents, scientific discovery and safety. After it you can reason about why deployed agents fail and design an evaluation harness that catches it.
Is Agentic AI MOOC (UC Berkeley, Fall 2025) free?
Agentic AI MOOC (UC Berkeley, Fall 2025) is free to access.
What level is Agentic AI MOOC (UC Berkeley, Fall 2025) for?
Agentic AI MOOC (UC Berkeley, Fall 2025) is aimed at a advanced audience. Recommended background: Working knowledge of large language models, prompting and tool calling, Comfort reading ML research papers — lectures cite them without recapping, Python and prior experience building at least one LLM application, for the project tiers.
How long does Agentic AI MOOC (UC Berkeley, Fall 2025) take?
Expect roughly 12 lectures x ~2h, Sep 15 - Dec 8 2025; free to audit on video, self-paced. Most learners work through it at their own pace.
What will I learn from Agentic AI MOOC (UC Berkeley, Fall 2025)?
You'll learn: Trace how agent system design evolved from single-call prompting to verifiable post-trained models; Evaluate agents properly, including what lab benchmarks systematically fail to measure in production; Compare multi-agent coordination approaches from two labs that built them at frontier scale; Understand post-training and reinforcement learning recipes used to make agentic models reliable; Apply agents to scientific discovery workflows including hypothesis generation and experiment design; Identify the latency, robustness, safety and user-trust failures that only surface after deployment; Reason about embodied and interactive agents where the environment, not text, is the interface; Assess agentic AI security threats and the defences currently available against them.
Topics
Sources
This page was written from 2 sources, 1 on domains other than agenticai-learning.org.