The ReAct loop, and createAgent

The ReAct loop, from first principles: call the model; if it requested tools, execute them and append the results as tool messages; call the model again; repeat until it stops requesting tools and returns a final answer. Reason, act, observe, repeat. That loop is all "agent" means.

import { createAgent, tool, toolStrategy } from "langchain";

const agent = createAgent({
  model,
  tools: [snapshotInbox, applyPlan],
  systemPrompt: "You triage the Todoist inbox. Emit file actions only.",
  responseFormat: toolStrategy([TriagePlanSchema]),
});
const out = await agent.invoke({ messages });
out.structuredResponse; // typed TriagePlan when responseFormat is set

responseFormat: toolStrategy([ZodSchema]) is how you get typed structured output out of the agent loop, surfaced as result.structuredResponse. That is the modern replacement for the bare withStructuredOutput call from module 1, when you are inside an agent rather than calling a model directly.

Middleware wraps the loop with cross-cutting behavior, composed in a middleware: [...] array: summarizationMiddleware compresses long histories automatically, humanInTheLoopMiddleware pauses on chosen tools for human sign-off (module 5's territory), piiRedactionMiddleware strips PII. For this module you want the loop plus structured output; the destructive-action gate comes in module 5.