Agentic AI Workshop 2026 - Session 1: Models, Tools & Agent Harnesses

What actually happens when an AI agent checks the time, uses a tool, remembers a conversation or takes an action? In the first session of the 2026 NYU Agentic AI Workshop, Adi Singhal breaks an agent into three parts: the model that supplies judgment, the harness that manages the work, and the software that executes tools. A restaurant analogy and a live request inspection make those boundaries concrete. The full session includes the class app store walkthrough, the first building assignment and audience Q&A. What you'll learn Separate the model, harness and tool execution using Lalima, Harry and Tim. Inspect system instructions, user input, tool requests, streaming events and tool results. Distinguish conversation history, memory, handoffs and provider-side prompt caching. Understand instruction priority and the role of untrusted tool output. Evaluate eight agent-design dimensions: control, trigger, authority, duration, environment, topology, memory and interface. Explore a small Python harness and a working demo, then publish an agent with a custom tool. Choose designs through product requirements and evaluation of speed, cost and accuracy. Chapters 00:00 Welcome and workshop format 03:02 Instructor introduction 07:15 Tools, inference and the goals for Session 1 09:03 Student project ideas and Demo Day 17:16 From chat and tool use to personal agents 22:07 The restaurant video 23:28 Lalima, Harry and Tim: model, harness and tools 30:52 Spotting the mistake in the analogy 35:05 Live demo: asking an agent for the time 41:27 Inspecting the request payload 44:35 Streaming responses and tool execution 47:19 Q&A: streaming and inference providers 50:36 Visible reasoning and model decisions 54:10 Who manages conversation history? 56:31 Prompt caching and repeated context 1:02:00 System prompts and instruction priority 1:05:40 Eight dimensions of an agent 1:06:11 Control: fixed steps or adaptive choices 1:07:09 Trigger: a person, schedule or event 1:11:51 Authority: acting or asking for approval 1:13:45 Duration: one task or an ongoing responsibility 1:15:57 Environment: tools and browsers 1:16:25 Topology: one agent or an orchestrator 1:17:45 Memory across interactions 1:18:40 Interface beyond chat 1:19:38 The small Python harness and starter code 1:20:43 Class app store walkthrough 1:21:39 Homework: build an agent with a custom tool 1:25:22 Launching apps and visitor-funded inference 1:27:06 Exploring the harness dimensions in the demo 1:29:22 Demo Day and coffee chats 1:30:08 Packaging, publishing and feedback 1:35:04 Open Q&A 1:36:01 Prompt caching versus memory and handoffs 1:37:16 Choosing designs and evaluating performance 1:42:16 Last year's workshop and alternative harnesses 1:44:21 External tools and MCP servers 1:46:12 Building an application versus using a coding agent 1:49:11 Can the harness call tools directly? 1:53:00 Where LangGraph fits 1:54:01 When a conventional pipeline may be enough 1:55:56 Model behavior, safety and harness controls 2:00:44 Keeping up with agentic AI Workshop resources Course page and Session 1 slides: https://adityasinghal.com/agentic-ai-... Public workshop repository and starter code: https://github.com/adityaarunsinghal/... Class app store for registered participants: https://nyu-workshop.adityasinghal.com/ First assignment Build a working variant of the starter or your own agent with at least one custom tool. Develop it locally, package it and upload it to the class app store before Session 2. Explain what the tool does and what feedback would help, then try other participants' apps. About the instructor Adi Singhal is an applied scientist at Amazon and an NYU alumnus who studied psychology and data science. His experience includes software engineering at Amazon Health and Amazon Bedrock. This is an independent educational workshop, not affiliated with or endorsed by Amazon or AWS. Remaining workshop dates Session 2: October 15, 2026 Session 3: October 22, 2026 Demo Day: October 29, 2026 Recorded at NYU on October 8, 2026. #AgenticAI #AIAgents #LLM #ToolCalling #MCP #NYU