Harness Engineering: Build an AI Coding Agent

About this course

Learn how AI coding agents really work by building one yourself.

Tools like Claude Code and Cursor can feel like magic. They are not. Underneath is a small, understandable idea called a harness: the code around a language model that lets it act, remember, stay safe and finish a job. In this course you build a complete harness from scratch, in plain Python, with no frameworks and no third-party packages.

Your agent is called Zill. You start with the core idea and end with a working command-line coding agent.

What you will build:

• The model client, with a single doorway to the API, retries and a neutral message format
• The agent loop, the heartbeat that asks the model, runs tools and feeds results back

Six real tools (read, write, edit, run commands, list, search) that run locally, in a sandbox or over SSHA permission gate with safe, read-only and yolo modes, plus a path jail that keeps the agent inside one folder

Context compaction, so long tasks fit inside a token budget

Long-term memory, skills and crash-proof sessions you can kill and resume

Sub-agents that take on a task with a clean context and report backA fleet runner for parallel agents, and a command-line front door

Why this course is different: every lecture adds one small file, explains the design rules behind it, and ends with a demo you can run and test. You see why each part exists, not just how to copy it.

Who it is for: developers who want to understand agents instead of treating them as a black box.

What you need: basic Python and a model API key. A free key is enough to follow along.

By the end you will have a working agent you understand line by line, and the confidence to extend it.

Prices, discounts and availability are set by Udemy. We may earn a commission when you purchase through links on this site.