Summary

This repository provides a set of production-ready design patterns for building reliable, safe, and scalable AI coding agents. Distilled from over 512,000 lines of Claude Code, these patterns address critical aspects of agent "harnessing" beyond the core LLM loop, such as memory, permissions, context management, and multi-agent coordination. It targets engineers developing agent runtimes, custom agents, or orchestration systems, offering a practical blueprint for robust AI agent development.

How It Works

The project systematically analyzes the Claude Code runtime to extract 6 core harness design patterns: Memory, Skills, Tools & Safety, Context Engineering, Multi-agent coordination, and Lifecycle. Each pattern is presented as a portable, runtime-agnostic principle, grounded by evidence from the Claude Code codebase but not tied to its specific implementation. This approach allows developers to apply these battle-tested concepts to diverse agent stacks, focusing on the surrounding infrastructure that ensures reliability and scalability.

Quick Start & Requirements

Installation is straightforward via the Vercel Skills CLI: npx skills add github:keli-wen/agentic-harness-patterns-skill. The skill is compatible with Claude Code, Codex, and over 40 other agents within the open agent skills ecosystem. No other specific prerequisites like GPUs or datasets are mentioned.

Highlighted Details

• Patterns derived from a deep, source-level analysis of 512,000 lines of Claude Code.
• Covers 6 fundamental harness layers with 11 detailed reference documents.
• Principles are designed to be portable and applicable across different agent runtimes.
• Roadmap includes analyzing Codex CLI and Gemini CLI to synthesize cross-runtime fundamental patterns.

Maintenance & Community

The project is maintained by keli-wen. While specific community channels like Discord or Slack are not listed, it integrates with the broader "open agent skills ecosystem." The roadmap indicates future expansion by analyzing other major agent runtimes.

Licensing & Compatibility

The project is licensed under the MIT license. While compatible with various agent systems, a disclaimer states the content is for "educational and learning purposes only" and does not reproduce proprietary source code, suggesting careful consideration for direct commercial integration beyond learning.

Limitations & Caveats

This resource focuses exclusively on design patterns distilled from Claude Code and is intended for educational purposes. It is not a user guide, prompt engineering tutorial, or a direct code dump. Analysis of other agent runtimes like Codex CLI and Gemini CLI is planned for future releases.