Notebooks

The tracks

Notebooks are organized into curated tracks in the left sidebar:

• Framework tracks — LangChain (and LangGraph), LlamaIndex.ts, AWS Strands Agents SDK, OpenAI Agents SDK, Vercel AI SDK, Google ADK, and Multi-Agent Systems with LangGraph.js. Each track teaches the framework's own primitives (agents, tools, handoffs, state machines) in its idiomatic style.
• Foundations Lab — prompt engineering and multimodal work built on raw fetch, so you see the wire format before any framework hides it.
• Agentic Evals — eight notebooks from deterministic checks through LLM-as-judge, juries, the RAG triad, trajectory grading, red-teaming, and operational metrics.
• Failure Modes Lab — one long notebook that reproduces twelve real production failures (runaway loops, prompt injection, RAG poisoning, context rot, cost blow-ups) and ships the fix for each.
• Standalone Agents, Enterprise Ops & Safety, Real-world Examples — applied patterns: cost routing, semantic chunking, MCP servers, and end-to-end builds.

How cells run

• Each code cell is a real editor (CodeMirror) — edit the code, then run it with the play button or Shift+Enter. Output, errors, and run duration appear under the cell.
• Cells execute in your browser. Framework packages (LangChain, LlamaIndex, and the rest) are loaded as version-pinned ES modules from a CDN at run time.
• Model calls go through an authenticated, OpenAI-compatible proxy scoped to the notebook (/api/notebooks/…/ai/v1 with chat, embeddings, and image endpoints), so notebooks work without putting an API key in the browser.
• A Reset control restores every cell to its original source and clears outputs — experiments are free.

Where notebooks fit

Notebooks teach the code-level view of the same patterns the rest of the platform teaches visually. A typical loop: meet a pattern in the lessons, run it as a graph on the Swarm Canvas, then open the matching notebook to build the same thing in framework code you could ship anywhere.