Harness Engineering 101 - Learning AI Agent Implementation in About 300 Lines of Python

What the book is

The book is a series of 16 chapters about AI agents. Its main idea is that an AI model only receives a list of messages and sends back text. The program that manages that list is called the harness. The author says most ideas called "agents" are small changes to this one loop. Each change fixes a real problem that came up.

Key idea: brain and body

Think of the LLM (large language model, an AI that predicts text) as the brain. Think of the harness as the body. The LLM API is stateless, which means it does not remember past turns. So the harness must send the whole conversation again every time, as a JSON array (a structured list written in a common data format).

Key facts

• The book has 16 chapters.
• It is split into four parts plus an epilogue and an appendix.
• The agent grows to about 300 lines of plain Python.
• Version 1.1 was published in October 2026.
• The author also works on One Code, an open-source coding agent.

Why it matters

Many people find AI agents confusing or magic. This book shows that they are built from simple parts. Students and developers can read the code, understand each step, and learn how to build their own agent.