Harness 101: Engineering the Loop Around the Model
Almost everything that looks like a different "advanced agent" - Deep Research, Plan-then-Act, coding agents, multi-agent systems - is the same ReAct loop with different restraints bolted on. This course peels that sentence apart one layer at a time: the loop, the tools (a tool is a contract, not a capability), context as a budget, memory that survives a reset, and the production scaffolding of skills, hooks, and prompt architecture. Every lesson is concept-first with an optional Build It appendix that assembles, across the course, one runnable agent in Python. Built for engineers who want to read, debug, and design any agent from first principles.
Created by
Jason Zhou · Founder, AI Builder Club
Course Outline
Want the theory first? Read our free AI Agents 2026 guide - read it anytime, no signup.
Module 0 – Orientation
What a harness is, and why it - not the model - is the product you own.
Module 1 – Foundations: The Loop
The ReAct loop as the one primitive, harness anatomy, Deep Research, and Plan-then-Act.
Module 2 – Tools: The Model's Hands
A tool is a contract written for the model, the quality of that contract, the built-in toolset, and the agent file system.
Module 3 – Context Engineering
Context as a metered budget, offloading to disk, compaction and microcompact, and progressive disclosure.
Module 4 – Memory & Knowledge
Why data is not memory, the three-layer memory model, and building a memory substrate that heals itself.
Module 5 – Orchestration: From One Run to Many
You are already a loop, open vs closed loops, ReAct to orchestration, and the loop's evolution.
Module 6 – Production: Extending & Shipping a Harness
Skills, hooks, and plugins; prompt architecture from a real system prompt; writing for agents; the frontier; and a capstone harness design.