A Chinese-language autopsy of how AI agents actually work

What it does

This repo distributes two things: a series of Chinese-language PDF deep-dives into production agent source code (ClaudeCode, Hermes Agent, memory systems), and a minimal Python agent skeleton for teaching purposes. The PDFs are the main attraction; the code is a deliberately stripped-down CLI demo.

The interesting bit

The author treats “agent” as an architectural pattern to dissect, not a product to sell. The teaching code uses a swappable Fake LLM so you can study the loop, skill discovery, and CLI wiring without burning API credits or drowning in framework abstraction.

Key highlights

• Chinese-language PDF reports analyzing real agent codebases (ClaudeCode, Hermes Agent)
• Minimal Python agent with Poetry-managed deps, ~3 files of core logic
• Fake LLM interface designed for drop-in replacement with real remote APIs
• Skills directory auto-discovery via CLI flag
• Version 2.1 adds a chapter on memory systems

Caveats

• The “deep dive” content is in PDFs, not the repo itself; the repo holds analysis materials, not full source mirrors
• The teaching agent is explicitly not production-ready: no real LLM API wired up yet
• Some newer reports (e.g., Hermes Agent) are gated behind a paid knowledge-planet community

Verdict

Worth a bookmark if you read Chinese and want architectural context before building your own agent. Skip it if you need a batteries-included framework or English-first documentation.