Nodes and edges

A graph runtime is smaller than its reputation. Visit a node, run it, update state, choose an outgoing edge, repeat until you reach an end node or run out of steps. The version powering this lab is about a hundred and twenty lines.

while (current && steps < maxSteps) {
  const node = graph.nodes.find((n) => n.id === current);
  await runNode(node, state);

  const outgoing = graph.edges.filter((e) => e.from === node.id);
  const next = node.kind === "router"
    ? outgoing.find((e) => matches(e.when, state)) ?? outgoing.find((e) => !e.when)
    : outgoing[0];

  current = next?.to ?? null;
}

Worth knowing before adopting a framework: most of what they add is ergonomics, typing and persistence rather than a different execution model. Understanding the loop above means you can read what a framework is doing and decide whether you need it.

Four node kinds cover most workflows. An llm node calls a model. A tool node calls a function. A router node reads state and picks an edge. An end node stops. This lab adds a fifth, interrupt, covered in lesson 5.