Why We Built TypeScript Notebooks (and Skipped Python)

You can read every tutorial ever written about AI agents and still not understand them. We know this because we did it, and then we watched learners do it. The concepts slide off. Tool calling, ReAct, handoffs, RAG, guardrails — they stay abstract until the moment you run one, change a line, and watch the behaviour move. So we stopped writing things to read and started shipping things to run.

AgentSwarms now has 50+ runnable TypeScript notebooks: real LangChain, LangGraph, the Vercel AI SDK, the OpenAI Agents SDK, and LlamaIndex.TS — not screenshots, not pseudo-code, but actual library code that executes when you click a button. No install. No kernel. No API key to paste. You open a notebook, press run, and an agent does something in your browser tab.

const swarm = ctx.lc.langgraph;        // real LangGraph, in your browser
const out = await researchSwarm.invoke({
  prompt: "Brief me on the EU AI Act timeline",
});
ctx.log(out.summary);
return out.meta;

Why runnable beats readable

A static code block is a photo of a meal. It tells you what the dish looked like; it doesn't feed you. The gap between seeing agent.invoke() and running it — then breaking it, then fixing it — is the entire gap between recognizing agentic AI and understanding it. Recognition is what you get from videos and docs. Understanding is what you get from your hands on a live cell.

How to actually learn from them

The notebooks reward an active reader, not a passive one. A few ways learners get the most out of them:

• Follow a track, in order. Start with LangChain fundamentals or standalone agents, then climb to multi-agent LangGraph. Each notebook assumes the last one clicked.
• Never run a cell without changing it. Read it, run it once to see the baseline, then change exactly one thing — the prompt, the model, a tool — and re-run. One variable at a time is how you learn what each piece does.
• Use them as a cookbook. Building a refund agent, a contract analyzer, a router? There's probably a real-world notebook that's 80% of your starting point. Copy the shape, keep the patterns.
• Pair them with the canvas and labs. Prototype the architecture visually, run the code version in a notebook, then break a swarm on purpose in the Failure-Mode Labs. Three angles on the same idea.
• Learn the TS agent ecosystem specifically. This is where you'll find LangGraph.js, the Vercel AI SDK, and the OpenAI Agents TS SDK taught hands-on — a stack almost no course covers.

“But where are the Python notebooks?”

Let's be straight about this, because it's the first question every experienced practitioner asks. Python is the lingua franca of machine learning, and most agent tutorials you've seen are Jupyter notebooks on Colab. We don't have those, and it's not an oversight — it's a consequence of where AgentSwarms runs.

AgentSwarms is a web platform that runs on edge infrastructure (Cloudflare Workers), with no per-user Python servers to spin up and no GPUs to rent. There's no Python runtime in that picture, and the browser can't run Python natively. The technically-possible workaround — Pyodide, Python compiled to WebAssembly — is heavy to load and, more importantly, still can't run the real agent frameworks people actually deploy. A 'Python notebook' on our platform would be a slow, fake Python that couldn't import the libraries that make Python worth using. That's a worse lie than not offering it.

Why TypeScript-in-the-browser turned out to be the better classroom

Here's the part we didn't expect: the constraint made the learning experience better, not worse. Running TypeScript directly in the browser removes the single biggest reason people bounce off tutorials — the setup.

• Zero setup. Nothing to install, no kernel to connect, no runtime queue, no pip install that breaks on a dependency conflict. The cell just runs.
• Instant feedback. Click, and output appears below the code in the same second. Tight loops are how skills form.
• It's where agents actually ship. A huge share of production agents run on the edge, in serverless functions, and in the browser — in JavaScript and TypeScript. LangChain.js, LangGraph.js, the Vercel AI SDK, and the OpenAI Agents TS SDK are first-class citizens here, not afterthoughts. You're learning the deployment language, not just a teaching language.
• Real libraries, not toys. The cells import genuine LangChain and LangGraph and run them live. You're reading and editing the same APIs you'd use in a real project.
• Your keys stay safe. You never paste a provider API key into a cell.

How this differs from the rest of agentic-AI education

There's no shortage of ways to learn about AI agents — video courses, Colab notebooks, framework docs. They're good at what they do. But for agentic AI specifically, they share the same blind spots, and that's where AgentSwarms is built differently.

• The TypeScript agent ecosystem, taught hands-on. LangGraph.js, the Vercel AI SDK, the OpenAI Agents SDK in TS — barely covered anywhere else, fully runnable here.
• Zero friction to the first run. No install means a beginner reaches a working agent in seconds, not after a setup tutorial they never finish.
• We teach the failure modes. A whole evals track and the Failure-Mode Labs exist precisely because every other course stops at the happy path — and the happy path is not where production lives.
• It's a platform, not a playlist. Notebooks sit beside a visual swarm canvas, a model playground, and interactive decks. You can read a concept, run it, build it, and break it without leaving the site.
• You can start for free, immediately. Real model calls run through a built-in gateway — no 'bring your own paid API key' wall before lesson one.

We didn't set out to make a statement about Python versus TypeScript. We set out to remove every excuse between a curious person and a running agent. Taking Python off the table forced us to lean all the way into in-browser, zero-install, production-relevant TypeScript — and that turned out to be exactly the classroom agentic AI needed. The best way to learn to build agents is to build them. Now you can, in the next tab.