A Thin Book on AI Agents
A technical book that demystifies AI agents with minimal code. Built on the insight that "the loop is simple — the hard part is handling data," it guides you through practical agent development.
What you'll learn
Agent Loop
Starting from a 14-line LLM call, build a complete agent in 72 lines. Understand the essence of the loop with minimal code.
Mini-SWE-agent
Analyze the code of a practical agent that achieved 74%+ on SWE-bench. The same core loop is used in production systems.
PageIndex (RAG)
A tree-structured RAG system that doesn't rely on vector search. Includes a case study using Toyota's earnings report.
Bonus Tracks
Source code analysis of Gemini CLI, Codex CLI, and OpenClaw. A deep dive into the design of real-world agents.
Table of Contents
- Agent Loop — From 14 Lines to 72
- Mini-SWE-agent — Anatomy of a Practical Agent
- Why Build Your Own Agent — Data Strategy
- Bonus: Reading the Gemini CLI Source
- Bonus: Reading the Codex CLI Source
- Bonus: PageIndex Implementation Guide
Tech Stack
- Python
- Gemini API
- Tool Calling
- Agent Loop
- RAG
- LuaLaTeX