CIS 4394 Agentic AI · Fall 2026 · Dr. Xinyu Fu
CIS 4394 Agentic AI · Fall 2026 · Dr. Xinyu Fu
Universities started teaching agents as a subject of its own in 2024, and most of the good courses publish their syllabi, slides and sometimes lectures. This page collects the ones with a public website, says what each emphasizes, and notes how this course differs. Every link was opened and checked on the date above; course sites move, so tell me if one breaks.
The shift from single models to compound AI systems: students build RAG, tool use and the agent loop from scratch, then study how frameworks such as DSPy abstract them. Distinctive lectures: data for agentic systems (traces, flywheels, synthetic data), optimization (prompt optimization vs weights), LLM-as-judge and evaluation infrastructure, and proactive agents. Two applied homeworks and a quarter-long project; the first homework is a public evaluation harness.
LLM-based agents that perceive, reason, plan and act over many steps: calling tools, navigating interfaces, writing code, working alongside people. Covers the foundations of building agents and the open research questions around long-horizon reasoning, training, safety and interaction, with an emphasis on hands-on implementation and on the training methods that make agents more capable and reliable. Ends with a poster session in December. Neubig co-founded All Hands AI, the company behind the OpenHands coding agent, so that material is first-hand.
The third iteration of Berkeley RDI's open course (after LLM Agents, Fall 2024, and Advanced LLM Agents, Spring 2025). A lecture series by people who build the systems: OpenAI, Google DeepMind, NVIDIA, Microsoft, Meta, Sierra and Sony AI. Recordings are on YouTube and anyone may watch and audit; assignments are optional, and a companion AgentX competition runs alongside. The best single place to hear the industry's own framing of the field, and the least structured as a course.
Follows one agentic system from the agent loop down through the serving stack: students first use a modern coding agent (Claude Code), then design their own agent with provided API access, measure it, and optimize the systems underneath. Five assignments with a leaderboard and a final project. Prerequisites are real: CS61, one graduate systems course, and comfort with Python and PyTorch.
The closest cousin to this course: a business school, no formal prerequisites, hands-on from day one. An MBA elective that also appears in the MS in Business Analytics and AI bulletin. Six three-hour sessions (building an agent; agent architecture; the power of agentic AI; enterprise agents and evaluation; deploying agents, with an industry panel; demo day). Students ship a working agent on an n8n backend with a front end, submit weekly one-page deliverables, and peer-review two demos. Public: the course site and the syllabus, with a student-projects page promised.
The information-systems angle on agents: ERP is the baseline, and the course is about designing, integrating and using business systems with AI-driven capabilities layered on top. Fourteen sessions in four modules, project-based: student teams build an incremental AI integration solution and present it in a final competition judged by industry partners. Open to all NYU graduate students, Stern included; a public syllabus is linked from the program page.
A studio course: teams of two or three build autonomous agents that plan, coordinate and execute complex workflows, covering agent design patterns, agent-to-agent communication, browser automation and decentralized coordination. Venture-style, with weekly mentor meetings, startup coaches and a Demo Day in May; the stated goal is an MVP at seed stage. Light on mechanism, heavy on shipping. Earlier iterations ran as AI Venture Studio and AI for Impact.
A hands-on graduate workshop in which much of the programming happens in class. It starts with small open-source models that run on a free Colab GPU or a laptop, then moves to hosted models and weight fine-tuning (QLoRA); one paper a week, a GitHub portfolio review, and a final project worth over half the grade. The closest in spirit to our local-first weeks. The same site publishes the interactive Transformer Logit Lens ↗ linked from our Week 9 page.
An applied course for graduate students and advanced undergraduates in engineering: LLM fundamentals, prompting, RAG and function calling, then agents for data analysis, simulation control, documentation and code synthesis, with examples from mechanical design, materials discovery and computational biology. Prior AI or machine-learning coursework and Python are expected.
Not an agents course in name, but a full public treatment of working with coding agents: MCP, agent skills, spec-driven development, "loop engineering" and the software factory, taught through weekly lectures, hands-on coding sessions and industry guests. The syllabus and the Fall 2025 archive are public. Pairs with our Week 2.
A project-based course (the project is 70% of the grade) on building and evaluating agentic applications on foundation models: background in foundation models, prompting and RAG, then agent architectures, with evaluation methods for each. Slides and homeworks are posted; lectures are not recorded. The Autumn 2024 offering is archived on the site.
Adjacent, not the same subject: MIT 6.7980 Topics in Multiagent Learning (Fall 2026, Daskalakis and Farina) studies multi-agent systems through game theory and learning theory rather than LLM agents; useful background for Week 9, not a substitute. Course site ↗
| Resource | What it is | Where we use it |
|---|---|---|
| Hugging Face Agents Course ↗ | Four units plus bonuses: agents and LLMs, frameworks (smolagents, LlamaIndex, LangGraph), agentic RAG, and a final project with automated evaluation; Apache-2.0. | Weeks 2, 5–8 |
| Microsoft · AI Agents for Beginners ↗ | Eighteen lessons with code, from tool use and agentic RAG to MCP and A2A, memory, local agents and securing agents; MIT licence. | Weeks 4–8, 10, 11 |
| LangChain Academy ↗ | Short courses on LangGraph, memory and no-code agents. | Weeks 2–4, 7 |
| DeepLearning.AI short courses ↗ | One- to two-hour builds on tools, RAG, MCP, multi-agent systems and evaluation. | Weeks 4–9 |
| Anthropic · Building Effective Agents ↗ | The essay most of these courses assign in week one, including ours. | Weeks 1, 2, 5 |
No systems sequence assumed. Group work has a code path (plain Python on a local model) and a no-code path (LangSmith Fleet), and the course asks the manager's questions first: who owns control flow, what does a run cost, when is it trusted unattended.
From Week 5 the agents run on a small model on your own laptop, so there is no API bill, no key to leak, and the unreliability of small models becomes the week's lesson instead of a surprise.
Evaluation (Week 8), security (Week 10) and agentic commerce (Week 11) are full weeks, and the capstone is graded on an eval harness and a security review, not only on whether the demo runs.