Video
Design, build, and deploy intelligent AI agents.
Study agentic AI with Stanford School of Engineering faculty: as a professional course you can take alongside your job, or as graduate coursework for academic credit.
Simple agents fall apart the moment the task gets messy. They can't check their own work, recover when something goes wrong, or carry context across sessions. Stanford's agentic AI courses teach you how agents do all three.
You'll work through end-to-end agent workflows: test-time scaling along with tool use, reasoning, planning, memory, and evaluation. The material is grounded in current Stanford research, and it pairs the theory with the engineering practice you need to get an agent into production. Where you start depends on how much time you have and what you want to walk away with. Some of these courses are self-paced and open now. Others follow the academic calendar and carry credit.
- Apply test-time scaling to get more out of a model at inference
- Explore methods for self-improvement and tool use
- Add retrieval and long-term memory to LLMs
- Build multi-step reasoning capabilities
- Design robust evaluation frameworks
Here's what you can expect...
Stanford Faculty
Learn from Stanford School of Engineering faculty who research agentic AI and teach it on campus.
Hands-On Assignments
Complete coding and system-building assignments. Coursework builds toward agents you can run and evaluate.
100% Online
Watch lectures, join live sessions, and work through problems alongside other learners.
Earn a Certificate
Earn academic credit toward a graduate certificate or a Stanford Certificate of Completion.
Who should enroll?
AI & ML Engineers
Machine Learning Engineers looking to build and improve AI agents, AI Engineers / AI Practitioners building agent-based systems, and AI Researchers and technical practitioners exploring agentic systems
Software Engineers & Developers
Software Engineers and technical professionals developing AI-powered applications and developers working with LLM APIs who want to build more advanced workflows
AI Practitioners
AI practitioners improving existing LLM applications using newer research ideas
Decision Makers Advancing Existing Systems
CTOs/VPs of Engineering at AI-forward companies who need to understand the technical depth of modern agentic systems
Lorem Ipsum Dolor sit Amet
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Donec nunc purus, mattis rutrum est ut, finibus fringilla mauris. Mauris arcu massa, hendrerit et lobortis eu, scelerisque a felis. Curabitur sed magna purus. Lorem ipsum nunc purus.
Graduate or Professional?
Time Commitment
Achievement
Classmate Interactions
Cost
Graduate Certificate Courses
90–120 hours per course
Earn up to 18 units of academic credit that may contribute to a certificate or a degree
Frequent collaboration with other students taking courses at the same time
$$$$
Professional Certificate Courses
6–13 hours per course
Earn a certificate and, in some cases, professional education units
Potential to connect with other participants through private social media groups
$$
Interested in Artificial Intelligence? Check out these online certificate programs:
Lorem ipsum dolor
Venenatis laoreet lacus
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Nullam varius mi vitae ornare dapibus. Fusce risus mauris, convallis in turpis id, venenatis laoreet lacus.
Lorem ipsum dolor
Venenatis laoreet lacus
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Nullam varius mi vitae ornare dapibus. Fusce risus mauris, convallis in turpis id, venenatis laoreet lacus.
Lorem ipsum dolor
Venenatis laoreet lacus
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Nullam varius mi vitae ornare dapibus. Fusce risus mauris, convallis in turpis id, venenatis laoreet lacus.
Lorem ipsum dolor
Venenatis laoreet lacus
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Nullam varius mi vitae ornare dapibus. Fusce risus mauris, convallis in turpis id, venenatis laoreet lacus.
Lorem ipsum dolor
Venenatis laoreet lacus
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Nullam varius mi vitae ornare dapibus. Fusce risus mauris, convallis in turpis id, venenatis laoreet lacus.
Lorem ipsum dolor
Venenatis laoreet lacus
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Nullam varius mi vitae ornare dapibus. Fusce risus mauris, convallis in turpis id, venenatis laoreet lacus.
Lorem ipsum dolor
Venenatis laoreet lacus
- list item 1
- list item 2
- list item 3
Lorem ipsum dolor
Venenatis laoreet lacus
- list item 1
- list item 2
- list item 3
Lorem ipsum dolor
Venenatis laoreet lacus
- list item 1
- list item 2
- list item 3
Lorem ipsum dolor
Venenatis laoreet lacus
- list item 1
- list item 2
- list item 3
Lorem ipsum dolor
Venenatis laoreet lacus
- list item 1
- list item 2
- list item 3
Instructors
Aakanksha Chowdhery
Adjunct Professor, Stanford Computer Science
Azalia Mirhoseini
Assistant Professor, Stanford Computer Science
Core Topics in Agentic AI
Depth and coverage vary by course
Test-Time Scaling
Test-time scaling spends extra compute at inference to improve a model's output. You'll look at how candidate answers are generated and selected, and how to weigh the added compute cost against the quality you get back.
Self-Improvement: Rewards & Verification
Verifiers check an agent's work before you trust it. Output rewards score the final answer while process rewards score the steps that led there, and the two pull training and inference in different directions. Ensembles of verifiers often outperform a single model.
Self-Improvement
An agent improves when it pairs its own reasoning with feedback from the environment, which is the idea behind the ReAct method. Reinforcement learning extends that further, so a model learns from execution feedback and verification signals rather than from labeled examples alone.
Memory and Retreaval
An agent's context window is finite, so anything it needs to remember across sessions has to live somewhere else. Approaches like MemGPT and Cartridges extend memory past that limit, and retrieval brings the right information back at the moment the agent needs it.
Agents for Software Engineering
Software engineering is where agents have advanced fastest, in part because code can be compiled, run, and tested, which gives an agent an unambiguous signal about whether it succeeded. That work now reaches into CUDA kernel development and compiler optimization.
Long Horizon Tasks and Agentic Evaluations
Evaluation gets harder as tasks stretch from minutes to hours, because the question shifts from whether an agent answered correctly to whether it finished at all. Agents struggle most with sourcing their own information and with problems that call for judgment rather than a specification.
