$ cat ~/map/learn-pi-agent.md

learn-pi-agent

A chapter-by-chapter rebuild of Pi's agent harness, from one LLM call to a working coding agent.

tutorials✓ Sep 4

## TL;DR

A bilingual study project that rebuilds the Pi agent harness in ordered chapters — from a minimal LLM call through tool calling, sessions, compaction, skills, and extensions to an assembled coding agent.

## Key points

• $Chapters evolve in order from a minimal LLM call through parallel tool execution, steering and follow-up, context transformation, the session tree, compaction, skills, and the extension runtime to a final assembled coding agent.
• $Each chapter continues the previous chapter's open questions while keeping an independent implementation and its own tests.
• $Written against Pi's Agent Core and Coding Agent sources, examining where each mechanism's complexity lives — MCP, subagents, permission prompts, plan mode, and todos are deliberately left out of Pi's default product and composed when needed.
• $Chinese-first with an English README; the online docs build (yiz-hhh.github.io/learn-pi-agent) mirrors the chapters but is currently written in Chinese.
• $Runs on npm workspaces: npm install at the repo root, then npm run build and npm test type-check and test every chapter.

## FAQ

## References

• README ↗

Last updated: Sep 4, 2026