hendrixer/harness-engineering

This repository contains the materials for a one-day hands-on workshop on building reliable AI agent systems in TypeScript and Node.js. The workshop focuses on the infrastructure layer that surrounds a language model, covering the patterns that keep agents durable, isolated, and coordinated rather than just teaching how to make a single API call and process the response. The course is organized into seven lessons, each on its own git branch. Lesson 1 introduces the basic agent loop. Lesson 2 covers durable execution: checkpointing the agent's progress so that a crash or a rate limit error causes the agent to resume from where it left off rather than starting over. Lesson 3 covers sandboxing, meaning running code that the agent generates in an isolated environment with timeouts so it cannot affect the rest of the system. Lessons 4 and 5 address memory management and orchestration, including how to route tasks to specialized agents and pass context between them. Lessons 6 and 7 cover hierarchical supervision (a coordinator that spawns parallel sub-agents and merges results) and human-in-the-loop approval flows that can pause for minutes or days waiting for a person to respond. Each lesson branch contains the completed solution for the previous lesson, so a participant who falls behind can check out the next branch to catch up. Lesson notes are written in Markdown and can be read directly in a code editor, opened as an Obsidian vault, or served locally as a documentation site using VitePress. The included development setup starts a local server and a browser-based inspector that shows a chat panel on the left and a live event stream on the right. Students add code in each lesson and can watch events flow through the system in real time. The server uses Express and OpenAI's API via the Vercel AI SDK. No deployment is needed, everything runs locally. Prerequisites are comfort with TypeScript and Node.js and a basic understanding of how LLM APIs and agent loops work.