Andrew Ng teaches interactive agents workshop

DeepLearning.AI just added a new short course for a very specific problem in AI product building: agents are getting smarter, but most of them still show up as a chat box. That’s fine for demos. It’s weak for real software. The new course, “Build Interactive Agents with Generative UI,” is about turning an agent into something that can actually render interface elements — charts, forms, cards, whiteboards, and shared canvases — while the user works alongside it. ### What’s actually new here? The release itself is new — DeepLearning.AI’s community forum listed the course announcement on May 6, 2026. The instructor is Atai Barkai, co-founder of CopilotKit, and the course is live on Learn.DeepLearning.AI as a beginner short course. It’s not an Andrew Ng-taught class in the usual sense. It’s a DeepLearning.AI launch in Andrew Ng’s course ecosystem, with Barkai doing the teaching. (learn.deeplearning.ai) ### What does “generative UI” mean? Basically, it means the agent doesn’t just answer with text. It can decide that the right response is a pie chart, a flight card, a form, or a collaborative canvas, then render that interface on demand. That is a pretty big shift in product feel. A text reply tells you something. A generated interface gives you something to do. (community.deeplearning.ai) ### What do you build in the course? The course description is unusually concrete. You connect an agent to a React frontend using CopilotKit and the AG-UI protocol, wire in LangChain, and build a fullstack app where the agent can render UI components dynamically. It also covers MCP Apps for hooking into third-party applications, and ends with a canvas-style app where the agent and user share live state instead of passing messages back and forth. (learn.deeplearning.ai) ### How long is it? Short. The course page lists 1 hour 28 minutes, 8 video lessons, 5 code examples, and 1 graded assignment. That tells you what kind of thing this is. It’s not a semester-long theory course. It’s a compact builder-oriented walkthrough meant to get a working pattern into your hands fast. ### What’s the framework angle? (learn.deeplearning.ai) The tooling stack matters because this is not being pitched as a one-off toy. The course uses CopilotKit and AG-UI, and the description says AG-UI has integrations across LangGraph, Google, AWS, Microsoft, and more. It also mentions A2UI — an open spec co-developed with Google — for declarative layout assembly. So the pitch is portability: learn one interaction model, then carry it across multiple agent ecosystems. ### Why does this matter now? Because a lot of “agent products” still feel like wrappers around an LLM. You type. It replies. Maybe it calls a tool. But real work software usually needs visible state, editable objects, and shared context. If an agent can create the interface it needs in the moment, the product stops feeling like a chatbot bolted onto an app and starts feeling like the app itself. That’s the real promise here. (learn.deeplearning.ai) ### Is this really about Claude-style apps? That’s a fair inference, but the broader point is bigger than any one model company. The course is aimed at building fullstack agent applications where the UI changes with the task. Claude, ChatGPT, and other assistants have pushed users to expect richer outputs. This course is about giving developers the patterns and plumbing to build that interaction style in their own products. (learn.deeplearning.ai) ### Bottom line? This launch matters less as “another AI course” and more as a signal about where agent products are heading. DeepLearning.AI is now teaching developers to build agents that don’t just talk — they render, coordinate, and share a workspace with the user. That’s a much more serious product pattern. (learn.deeplearning.ai)