Everyone agrees you should be building agents. Nobody agrees on how. One camp says drag nodes on a canvas and ship this afternoon. The other says real agents live in code, in your repo, behind your own API. They're both right, and the argument is a distraction: the loop is the same either way. A model, a set of tools, a memory, and a stopping condition. Once you see that, the question stops being "which side is right" and becomes "which lane fits this job."

This playbook walks both lanes properly. The first half builds agents in n8n: the AI Agent node, tools it can actually call (including MCP servers), memory and RAG on the canvas, human approval gates, multi-agent patterns, and the Evaluations feature that tells you whether any of it works. The second half builds the same ideas in TypeScript with the Vercel AI SDK inside an Astro site: a streaming chat endpoint, real tool definitions with schemas, the ToolLoopAgent, approval gates in code, structured output, and MCP as the bridge that lets your n8n workflows and your code agents share the same tools.

It pairs with Mastering n8n (which covers hosting and hardening the platform itself) and Roll Your Own Coding Agent (which builds the raw loop from nothing), and once you're shipping agents, QA in the Era of AI shows what happens when you point them at your test suite. This one is about shipping: picking a lane, building the agent, and knowing when to switch lanes as the job outgrows the canvas.

This is a living document and will be updated as the tools and patterns evolve.