cellinlab/how-pi-agent-works

This repository is a Chinese-language tutorial that explains how a type of AI agent called Pi Agent works and walks you through building a simplified version of one yourself. The tutorial is written as a documentation website using a tool called VitePress, which turns Markdown files into a browsable site. You run a command to start a local web server, then read through the tutorial in your browser. The content is organized around learning, not raw source code. It covers core concepts such as the agent loop (the repeating cycle an AI agent uses to receive a message, decide what to do, call a tool, and respond), how messages are structured, how tool calls work, how conversation history is stored as a tree, and how context is compressed when a conversation grows too long. It also walks through the layers of the actual Pi Agent source code, explaining how the pieces connect. Alongside the written tutorial there are four TypeScript demo programs that build on each other. The first one is the smallest possible working agent loop. Each subsequent demo adds one more capability: tools, conversation state, and context compression. A fifth optional demo lets you point the code at a real AI API to run a live test. The final piece is a full teaching project built with React on the frontend and Node.js on the backend. Running it starts both a local web interface and an API server, giving you a working agent you built yourself to experiment with. The tutorial and all code are MIT licensed. The README is written in Chinese, and the project appears to be maintained by a single developer who also links to social media and a subscription community.