# Build an agent you can explain. Source: https://adityasinghal.com/agentic-ai-workshop/ [Adi Singhal / NYU CDS](https://adityasinghal.com/) NYU CDS · FALL 2026 Agentic AI Workshop Watch an agent choose a tool and see what runs it. Then build one around a problem you care about! **October 8, 15, 22 & 29, 2026** Thursdays, 5:30–7:00 PM ET CDS 7th Floor Open Space [Watch the recordings](https://adityasinghal.com/agentic-ai-workshop/#recordings) Registration has closed. [Watch Session 1](https://adityasinghal.com/agentic-ai-workshop/#session-1-recording) [Looking for last year’s workshop? Explore 2025](https://adityasinghal.com/agentic-ai-workshop-2025/) A project you could build Illustrative THE MORNING EDITION YOUR TOPICS. YOUR PACE. ## A briefing for your commute. You choose the interests. Your agent brings back the reading. READING PREFERENCE **Practical agent engineering** 01 ### A tool worth understanding A short explanation, why it fits your interests, and the original source. 02 ### A claim worth checking The evidence behind it, with uncertainty kept visible. ### What would the agent do? 1. Read your saved preferences. 2. Find sources and select useful material. 3. Prepare a briefing you can inspect. 4. Check its claims against the evidence. This is an example of a possible project. You will choose your own. In person at NYU CDS Slides, live demos & discussion OpenRouter inference provided THE COURSE ## Build your own agent over four Thursdays. From the first tool call to a project you can share. Sessions 1–3 are required. Demo Day is highly encouraged, even if you don’t present. 1. OCT **08** SESSION 1 Required ### Agents and harnesses What an agent is, when to use one, and the machinery that runs its tools, context, and decisions. - Agent loops - Tools & MCP - Harnesses [Watch the recording](https://adityasinghal.com/agentic-ai-workshop/#session-1-recording) THURSDAY 5:30–7 PM ET 2. OCT **15** SESSION 2 Required ### Agents beyond chat Give an agent an interactive surface. Explore generated interfaces, typed decisions, and experiments with feedback. - A2UI & MCP Apps - Jev - Auto-Research [Watch the recording](https://adityasinghal.com/agentic-ai-workshop/#session-2-recording) THURSDAY 5:30–7 PM ET 3. OCT **22** SESSION 3 Required ### Multi-agent systems Decide when to divide work, what each agent needs to know, and how to coordinate handoffs and review. - Context management - Coordination protocols - Evaluation THURSDAY 5:30–7 PM ET 4. OCT **29** SESSION 4 Highly encouraged ### Demo Day Show what you built. Discuss your project with classmates and industry guests, and learn from their questions. - Student demos - Industry feedback - Coffee chats THURSDAY 5:30–7 PM ET HOW WE LEARN ## Follow the agent’s decisions in class. Keep your laptop closed and ask questions as we predict what happens next. Build your own between sessions. [Slides and starter code](https://github.com/adityaarunsinghal/agentic-ai-workshop-2026) support your project. You’ll be encouraged to share your agent through the class app store. IN THE ROOM ### 60 minutes of material + 30 minutes of Q&A Technical storytelling through slides, live demonstrations, and conversation. BETWEEN SESSIONS ### About 1–2 hours of development each week Start with a research question, a recurring task, or a personal workflow. Make a small version that works, then develop it each week. SUPPORT ### Friday office hours Bring a broken tool call, a confusing code path, or an idea that feels too big. We’ll help you find the next step. Joining details will be shared with registered students. YOUR BUILDING MATERIALS ## Starter materials for your own project. Slides and starter code for Sessions 1–3 are available in the [public course repository](https://github.com/adityaarunsinghal/agentic-ai-workshop-2026). Each session’s materials will be released the day before class. 01 ### 2026 public repository The home for this year’s course materials. [Open the repository](https://github.com/adityaarunsinghal/agentic-ai-workshop-2026) 02 ### Session 1: Agents and harnesses Download the slides and explore the starter code for your own agent. [Download Session 1 slides](https://raw.githubusercontent.com/adityaarunsinghal/agentic-ai-workshop-2026/refs/heads/main/session-01-agent-harnesses/session-01-agent-harnesses-2026-10-07.html) [Open starter code](https://github.com/adityaarunsinghal/agentic-ai-workshop-2026/tree/main/session-01-agent-harnesses/demo-app) 03 ### Session 2: Agents beyond chat Explore the slides and demo app, then build an interactive agent. [Download Session 2 slides](https://raw.githubusercontent.com/adityaarunsinghal/agentic-ai-workshop-2026/refs/heads/main/session-02-beyond-chat-interfaces/slides/session-02-beyond-chat-interfaces-2026-10-15.html) [Open starter code](https://github.com/adityaarunsinghal/agentic-ai-workshop-2026/tree/main/session-02-beyond-chat-interfaces/demo-app) 04 ### Session 3: Multi-agent systems Explore the slides and compare single-agent, graph and swarm fact-checking. [Download Session 3 slides](https://raw.githubusercontent.com/adityaarunsinghal/agentic-ai-workshop-2026/refs/heads/main/session-03-multi-agent-topologies/slides/session-03-multi-agent-topologies-2026-10-22.html) [Open starter code](https://github.com/adityaarunsinghal/agentic-ai-workshop-2026/tree/main/session-03-multi-agent-topologies/demo-app) 05 ### Build and publish your agent Follow the setup, ZIP upload, peer feedback, and course-credit instructions. [Read the project guide](https://github.com/adityaarunsinghal/agentic-ai-workshop-2026#clone-and-build-your-own-agent) 06 ### Class app store Publish your project and explore what others in the workshop build. [Open the class app store](https://nyu-workshop.adityasinghal.com/) _(For registered students only)_ RECORDED IN CLASS ## Session recordings Watch the full session or find a topic in the chapter list. SESSION 1 October 8, 2026 ### Models, Tools & Agent Harnesses [View on YouTube](https://www.youtube.com/watch?v=UtjOl-LPcCk) Full session · 2:02:32 #### Find a moment Chapter links open YouTube at that moment. 43 chapters 1. [00:00 Welcome and workshop format](https://www.youtube.com/watch?v=UtjOl-LPcCk) 2. [03:02 Instructor introduction](https://www.youtube.com/watch?v=UtjOl-LPcCk&t=182s) 3. [07:15 Tools, inference and the goals for Session 1](https://www.youtube.com/watch?v=UtjOl-LPcCk&t=435s) 4. [09:03 Student project ideas and Demo Day](https://www.youtube.com/watch?v=UtjOl-LPcCk&t=543s) 5. [17:16 From chat and tool use to personal agents](https://www.youtube.com/watch?v=UtjOl-LPcCk&t=1036s) 6. [22:07 The restaurant video](https://www.youtube.com/watch?v=UtjOl-LPcCk&t=1327s) 7. [23:28 Lalima, Harry and Tim: model, harness and tools](https://www.youtube.com/watch?v=UtjOl-LPcCk&t=1408s) 8. [30:52 Spotting the mistake in the analogy](https://www.youtube.com/watch?v=UtjOl-LPcCk&t=1852s) 9. [35:05 Live demo: asking an agent for the time](https://www.youtube.com/watch?v=UtjOl-LPcCk&t=2105s) 10. [41:27 Inspecting the request payload](https://www.youtube.com/watch?v=UtjOl-LPcCk&t=2487s) 11. [44:35 Streaming responses and tool execution](https://www.youtube.com/watch?v=UtjOl-LPcCk&t=2675s) 12. [47:19 Q&A: streaming and inference providers](https://www.youtube.com/watch?v=UtjOl-LPcCk&t=2839s) 13. [50:36 Visible reasoning and model decisions](https://www.youtube.com/watch?v=UtjOl-LPcCk&t=3036s) 14. [54:10 Who manages conversation history?](https://www.youtube.com/watch?v=UtjOl-LPcCk&t=3250s) 15. [56:31 Prompt caching and repeated context](https://www.youtube.com/watch?v=UtjOl-LPcCk&t=3391s) 16. [1:02:00 System prompts and instruction priority](https://www.youtube.com/watch?v=UtjOl-LPcCk&t=3720s) 17. [1:05:40 Eight dimensions of an agent](https://www.youtube.com/watch?v=UtjOl-LPcCk&t=3940s) 18. [1:06:11 Control: fixed steps or adaptive choices](https://www.youtube.com/watch?v=UtjOl-LPcCk&t=3971s) 19. [1:07:09 Trigger: a person, schedule or event](https://www.youtube.com/watch?v=UtjOl-LPcCk&t=4029s) 20. [1:11:51 Authority: acting or asking for approval](https://www.youtube.com/watch?v=UtjOl-LPcCk&t=4311s) 21. [1:13:45 Duration: one task or an ongoing responsibility](https://www.youtube.com/watch?v=UtjOl-LPcCk&t=4425s) 22. [1:15:57 Environment: tools and browsers](https://www.youtube.com/watch?v=UtjOl-LPcCk&t=4557s) 23. [1:16:25 Topology: one agent or an orchestrator](https://www.youtube.com/watch?v=UtjOl-LPcCk&t=4585s) 24. [1:17:45 Memory across interactions](https://www.youtube.com/watch?v=UtjOl-LPcCk&t=4665s) 25. [1:18:40 Interface beyond chat](https://www.youtube.com/watch?v=UtjOl-LPcCk&t=4720s) 26. [1:19:38 The small Python harness and starter code](https://www.youtube.com/watch?v=UtjOl-LPcCk&t=4778s) 27. [1:20:43 Class app store walkthrough](https://www.youtube.com/watch?v=UtjOl-LPcCk&t=4843s) 28. [1:21:39 Homework: build an agent with a custom tool](https://www.youtube.com/watch?v=UtjOl-LPcCk&t=4899s) 29. [1:25:22 Launching apps and visitor-funded inference](https://www.youtube.com/watch?v=UtjOl-LPcCk&t=5122s) 30. [1:27:06 Exploring the harness dimensions in the demo](https://www.youtube.com/watch?v=UtjOl-LPcCk&t=5226s) 31. [1:29:22 Demo Day and coffee chats](https://www.youtube.com/watch?v=UtjOl-LPcCk&t=5362s) 32. [1:30:08 Packaging, publishing and feedback](https://www.youtube.com/watch?v=UtjOl-LPcCk&t=5408s) 33. [1:35:04 Open Q&A](https://www.youtube.com/watch?v=UtjOl-LPcCk&t=5704s) 34. [1:36:01 Prompt caching versus memory and handoffs](https://www.youtube.com/watch?v=UtjOl-LPcCk&t=5761s) 35. [1:37:16 Choosing designs and evaluating performance](https://www.youtube.com/watch?v=UtjOl-LPcCk&t=5836s) 36. [1:42:16 Last year's workshop and alternative harnesses](https://www.youtube.com/watch?v=UtjOl-LPcCk&t=6136s) 37. [1:44:21 External tools and MCP servers](https://www.youtube.com/watch?v=UtjOl-LPcCk&t=6261s) 38. [1:46:12 Building an application versus using a coding agent](https://www.youtube.com/watch?v=UtjOl-LPcCk&t=6372s) 39. [1:49:11 Can the harness call tools directly?](https://www.youtube.com/watch?v=UtjOl-LPcCk&t=6551s) 40. [1:53:00 Where LangGraph fits](https://www.youtube.com/watch?v=UtjOl-LPcCk&t=6780s) 41. [1:54:01 When a conventional pipeline may be enough](https://www.youtube.com/watch?v=UtjOl-LPcCk&t=6841s) 42. [1:55:56 Model behavior, safety and harness controls](https://www.youtube.com/watch?v=UtjOl-LPcCk&t=6956s) 43. [2:00:44 Keeping up with agentic AI](https://www.youtube.com/watch?v=UtjOl-LPcCk&t=7244s) SESSION 2 October 15, 2026 ### Beyond Chat, MCP Apps & Decision APIs [View on YouTube](https://www.youtube.com/watch?v=CnrdDizVvV8) Full session · 1:30:03 #### Find a moment Chapter links open YouTube at that moment. 33 chapters 1. [00:00 Welcome and recap](https://www.youtube.com/watch?v=CnrdDizVvV8) 2. [04:06 Agent design choices](https://www.youtube.com/watch?v=CnrdDizVvV8&t=246s) 3. [06:25 The model request bubble](https://www.youtube.com/watch?v=CnrdDizVvV8&t=385s) 4. [09:45 Chat agent demonstration](https://www.youtube.com/watch?v=CnrdDizVvV8&t=585s) 5. [15:21 What the harness does](https://www.youtube.com/watch?v=CnrdDizVvV8&t=921s) 6. [20:25 Why tools require another request](https://www.youtube.com/watch?v=CnrdDizVvV8&t=1225s) 7. [23:22 The newspaper demonstration](https://www.youtube.com/watch?v=CnrdDizVvV8&t=1402s) 8. [28:41 MCP tools and resources](https://www.youtube.com/watch?v=CnrdDizVvV8&t=1721s) 9. [30:05 Interactive newspaper layout](https://www.youtube.com/watch?v=CnrdDizVvV8&t=1805s) 10. [32:41 MCP Apps and extensions](https://www.youtube.com/watch?v=CnrdDizVvV8&t=1961s) 11. [34:53 The cost of formatting decisions](https://www.youtube.com/watch?v=CnrdDizVvV8&t=2093s) 12. [39:02 Decision APIs](https://www.youtube.com/watch?v=CnrdDizVvV8&t=2342s) 13. [42:02 Bag of words and embeddings](https://www.youtube.com/watch?v=CnrdDizVvV8&t=2522s) 14. [45:44 Transformer classifiers](https://www.youtube.com/watch?v=CnrdDizVvV8&t=2744s) 15. [47:28 Text classification and confidence](https://www.youtube.com/watch?v=CnrdDizVvV8&t=2848s) 16. [50:39 Calibration](https://www.youtube.com/watch?v=CnrdDizVvV8&t=3039s) 17. [53:32 Decision-model limits](https://www.youtube.com/watch?v=CnrdDizVvV8&t=3212s) 18. [57:13 Jev in the newspaper](https://www.youtube.com/watch?v=CnrdDizVvV8&t=3433s) 19. [57:52 Noul and factual verification](https://www.youtube.com/watch?v=CnrdDizVvV8&t=3472s) 20. [1:00:24 Research and revision loops](https://www.youtube.com/watch?v=CnrdDizVvV8&t=3624s) 21. [1:02:45 Autoresearch](https://www.youtube.com/watch?v=CnrdDizVvV8&t=3765s) 22. [1:06:13 AlphaEvolve](https://www.youtube.com/watch?v=CnrdDizVvV8&t=3973s) 23. [1:07:15 Recursive self-improvement](https://www.youtube.com/watch?v=CnrdDizVvV8&t=4035s) 24. [1:10:02 Open discussion](https://www.youtube.com/watch?v=CnrdDizVvV8&t=4202s) 25. [1:12:52 Classification probabilities](https://www.youtube.com/watch?v=CnrdDizVvV8&t=4372s) 26. [1:14:51 Domain-specific calibration](https://www.youtube.com/watch?v=CnrdDizVvV8&t=4491s) 27. [1:16:56 Consistent classification and caching](https://www.youtube.com/watch?v=CnrdDizVvV8&t=4616s) 28. [1:17:40 Local development and publishing](https://www.youtube.com/watch?v=CnrdDizVvV8&t=4660s) 29. [1:19:01 Jev, autoresearch and RSI](https://www.youtube.com/watch?v=CnrdDizVvV8&t=4741s) 30. [1:19:57 Changing open model behavior](https://www.youtube.com/watch?v=CnrdDizVvV8&t=4797s) 31. [1:22:51 MCP connections and extensions](https://www.youtube.com/watch?v=CnrdDizVvV8&t=4971s) 32. [1:25:18 Classification heads and text generation](https://www.youtube.com/watch?v=CnrdDizVvV8&t=5118s) 33. [1:28:32 The cost of autoresearch](https://www.youtube.com/watch?v=CnrdDizVvV8&t=5312s) OCTOBER **29** DEMO DAY SHOW WHAT YOU BUILT ## Meet other builders through your project. Share your work with classmates and industry professionals from OpenRouter, Amazon, Google, and startups. Judges will select students whose demos they like for coffee chats. Come to show your project, give feedback, or see what everyone made. FROM THE 2025 CLASS ## Get a feel for the room. Last year, we built a personalized newspaper with MCP. These highlights offer a look at the examples, questions, and discussions. [2025 highlights playlist](https://www.youtube.com/playlist?list=PLgF7i4LH-YxYvhXK-yywN7eFRjRjQrq89)[2025 course code](https://github.com/adityaarunsinghal/agentic-ai-workshop-2025)[2025 workshop website](https://adityasinghal.com/agentic-ai-workshop-2025/) [2025 RECORDING ### The restaurant analogy A familiar way into the roles behind MCP. Watch highlight](https://www.youtube.com/watch?v=LoTpDu_rSHE)[2025 RECORDING ### Agents beyond chat See how an agent can fit into a different interface. Watch highlight](https://www.youtube.com/watch?v=VCDpMWg78iw)[2025 RECORDING ### When to split the work Explore the decision to use multiple agents. Watch highlight](https://www.youtube.com/watch?v=U5BMtOLvcTI) YOUR INSTRUCTORS ## Meet your instructors. ![Adi Singhal waving hello](https://adityasinghal.com/workshops/2026/instructors/adi-wave-2026-10-09.jpg) ### Adi Singhal Applied Scientist at Amazon, former AWS Bedrock engineer, Ambassador at the Agentic AI Foundation (AAIF) and Instructor at NYU. At NYU, he studied psychology and data science, conducted research with faculty, and founded and led the Data Science Club to 1k members. [Connect on LinkedIn](https://www.linkedin.com/in/adi-singhal/) ![Luca Chang waving hello](https://adityasinghal.com/workshops/2026/instructors/luca-wave-2026-10-09.jpg) ### Luca Chang Works in the AWS Agentic AI organization and is an MCP specification maintainer and SDK contributor. Expert in building production-ready agentic systems. [Connect on LinkedIn](https://www.linkedin.com/in/luca-chang/) ## Before you join ### Who is the workshop for? Members of the CDS community, including students and alumni. Undergraduates should have completed DS-UA 112. Attend the first three sessions and plan for weekly project work. ### Where do we meet? CDS 7th Floor Open Space, 60 5th Ave, between 12th and 13th Streets, New York City. Thursdays, 5:30–7:00 PM ET. The workshop is in person. Contact the organizer about individual attendance questions. ### What will I need? Time to build between sessions and curiosity in the room. Find slides and starter code in the [public course repository](https://github.com/adityaarunsinghal/agentic-ai-workshop-2026). Registered students will receive inference access through OpenRouter. ### Is registration still open? Registration has closed. You can view the [original registration form](https://docs.google.com/forms/d/e/1FAIpQLSd_lhxUjjw0NMCvsf5qojymmS8Wur0tA26rqOnSCZarhkV31g/viewform) for reference. The form requires Google sign-in. [Contact the workshop organizer](mailto:tb116@nyu.edu) about form access, attendance, or other questions. FROM THE WORKSHOP FORMS ## Who signed up? Fictional voices based on replies as of October 6, 2026. Approximate counts overlap. 1. ### ≈65 Green Finding an idea Read my perspective I'm still choosing my idea. I'm excited to discover what an agent could do and get inspiration from other people's projects. I worry about picking something so broad that I never finish it. Help me find one small problem I care enough about to keep exploring. 2. ### ≈35 Purple Career momentum Read my perspective I know I want a useful project for a career conversation, though the topic is still open. I'm excited to build something worth showing and meet people who can give me feedback. I worry about ending up with a generic demo while interviews, classes and the presentation compete for my time. 3. ### ≈35 Orange Everyday problems Read my perspective I have an everyday annoyance in mind, like choosing a café, planning a day or keeping records up to date. I'm excited to make something I would use myself. I worry about wrong details and an idea that grows into an assistant for everything. One task made easier would be a good start. 4. ### ≈35 Yellow Workflows that last Read my perspective I have a work process or a project already in motion. I'm excited to finish a useful part that saves repeated effort. I worry about missing information, hidden mistakes and who approves the next action. Help me make a small promise I can keep when another person starts relying on the result. 5. ### ≈30 Pink Research questions Read my perspective I have a field or a research question in mind, but the experiment may still be open. I'm excited to explore simulations or connect evidence across a collection of material. I worry that a convincing story could outrun the evidence. I want a result I can inspect and a clearer question to pursue. 6. ### ≈25 Blue Care and access Read my perspective I know whom I want to help: someone waiting for care, learning support or a service they struggle to access. My idea might still be a large ambition. I'm excited by the possibility of making help easier to get. I worry about privacy, unfairness and mistakes, so people need a way to question the result. NYU CDS · AGENTIC AI WORKSHOP 2026 ## Bring your curiosity and build an agent of your own. Registration has closed. [Watch the recordings](https://adityasinghal.com/agentic-ai-workshop/#recordings)