Pi-Agent's missing manual: a book, a lab, and an AI skill

What it does dg-ai-notes is an open-source tutorial series that reverse-engineers Pi-Agent, a production-grade AI Agent SDK that serves as the runtime base for products like OpenClaw. It offers two parallel tracks: a ten-chapter source-code deep dive and a seven-chapter practical course where you build a deployable DataAgent from scratch. The whole package is presented in both TypeScript and Python, and doubles as a web book, a PDF, and a skill file for AI assistants.

The interesting bit The author didn’t just write documentation—he packaged the expertise into dg-piagent, a skill that drops into Claude Code, ZCode, Codex, or Pi itself so your agent can reference SDK patterns while writing code. That turns a static tutorial into a living coding partner, which is arguably the most honest use of “AI-native learning” I’ve seen lately.

Key highlights

• Ten-chapter source dive covering the Agent Loop, tool system, event-driven architecture, and session management, with 30+ diagrams
• Seven-chapter practical build (DataAgent) with runnable L01–L07 code samples meant to go from zero to deployed
• Dual TypeScript and Python editions, plus an interactive Agent Loop notebook for parameter tweaking
• dg-piagent skill maintained against pi-coding-agent v0.83.0, designed to generate configs, scaffold features, and debug tool errors via AI assistant context
• Web reader with dark/light themes, plus Markdown and PDF downloads for offline reading

Caveats

• All content is in Chinese, so non-readers will need translation tools
• The skill and tutorials are pinned to a specific Pi-Agent version (v0.83.0), so expect drift as the upstream SDK evolves
• This is a learning resource and skill pack, not the Pi-Agent framework itself—head to the upstream repo for the actual SDK

Verdict Worth bookmarking if you’re trying to understand how production agent SDKs like Claude Code or Cursor work under the hood, or if you’re actively building on Pi-Agent. Skip it if you just want a drop-in agent library without the computer-science homework.