Add comprehensive agent architecture pattern documentation

[zhibao-labs]

Summary

Added a comprehensive Chinese-language guide documenting the agent architecture patterns and best practices extracted from the learn-claude-code curriculum. This document serves as a practical reference for building intelligent agents across any domain.

Key Changes

• Core Philosophy: Establishes the foundational principle that "the model itself is the agent" and code should only provide tools, inject knowledge, and protect context
• Agent Loop Foundation: Documents the immutable core loop pattern used across all 12 lessons, with the stop_reason != "tool_use" exit condition as the single source of truth
• Progressive Layering Paradigm: Outlines 6 levels of capability stacking:
  • Level 0: Minimal viable agent (model + 1 tool + loop)
  • Level 1: Multi-tool dispatch mapping
  • Level 2: Planning and progress tracking with TodoManager
  • Level 3: Subagent isolation for context separation
  • Level 4: On-demand knowledge injection via Skills
  • Level 5: Three-tier context compression strategy
• Production Extensions: Covers advanced patterns including persistent task systems, background task management, and multi-agent team coordination
• Anti-patterns Guide: Provides a reference table of common mistakes and their solutions
• Minimal Reproducible Template: Includes a complete, runnable 50-line agent implementation as a starting point
• Implementation Checklist: Breaks down the journey from zero to production into 4 phases with concrete milestones

Notable Details

• Emphasizes "don't over-engineer" philosophy with concrete examples of when to add features
• Uses consistent Python code examples throughout for clarity
• Includes YAML/Markdown skill file structure for knowledge organization
• Documents the Anthropic API convention of appending tool results as "user" messages
• Provides practical guidance on tool count (start with 3-5) and context management strategies

https://claude.ai/code/session_01Nn3nKUcudz6NmFwTXhxHVZ

[vercel]

@claude is attempting to deploy a commit to the crazyboym's projects Team on Vercel.

A member of the Team first needs to authorize it.