Coding agents are not just models connected to a terminal. They are harnesses: runtimes that manage the agent loop, tools, context, sessions, memory, approvals, and sandboxes. This book shows how to investigate those layers from public source code and reproducible behavior.
You will learn a six-part framework, rerun one jsonstats task across five representative harnesses, read benchmark evidence without confusing model skill with harness effects, and choose tools by architecture rather than popularity. The book examines Pi, OpenCode, DeepSeek Harness, Claude Code, Codex, and the orchestration patterns around them.
The result is a practical source-level guide for developers who need to understand what an agent can do, why it behaves that way, and which workflow costs will remain when the model or provider changes.
