how-claude-code-works

Why we included this project

This study grew out of a practical problem: the authors wanted to understand a production coding agent, and the only way in was a 500K-line TypeScript codebase. Rather than skim it, they read the source with Claude Code's help and wrote down what they learned, so the chapters are grounded in real internals instead of general advice about how agents ought to work. You get the agent loop, the four-level context compression, the seven layers that guard shell execution, and how the tools coordinate through one common interface. For anyone building or maintaining AI agents, this reads like a field guide to the problems that show up once a demo becomes a deployed system. It is documentation rather than software, so you are not meant to run it, but you can read it online and keep it open while you inspect the code yourself, which makes it a reasonable starting point before you build your own.