shareAI-lab

learn-claude-code

// summary

This repository provides a comprehensive educational framework for building agent harnesses, which are the essential environments that allow AI models to perceive and act. It argues that true agency is learned by models during training, while the developer's role is to construct the tools, knowledge, and context management systems that enable these models to function. Through twelve progressive sessions, users learn to build robust, scalable agent architectures by reverse-engineering the principles behind Claude Code.

// technical analysis

This project promotes a 'harness-first' design philosophy, arguing that true agency is derived from trained models rather than complex procedural code orchestration. It addresses the common industry pitfall of 'prompt plumbing' by providing a structured framework for building the environment—or harness—that allows intelligent models to perceive, reason, and act effectively. By separating the intelligence (the model) from the infrastructure (the harness), the project provides a modular, scalable approach to building agents that can be applied across diverse domains beyond software engineering.

// key highlights

Provides a clear architectural distinction between the agent model and the supporting harness infrastructure.

Features 12 progressive learning sessions that incrementally build a sophisticated agent system from a basic loop.

Implements a modular tool-dispatch system that allows for easy expansion of agent capabilities.

Demonstrates advanced context management techniques, including subagent isolation and multi-layer context compression.

Includes a robust task-management system with dependency graphs and persistent file-based storage.

Showcases a multi-agent coordination pattern using asynchronous JSONL mailboxes for team-based workflows.

// use cases

Building modular agent harnesses with tools, knowledge, and context management

Implementing persistent task systems and subagent coordination for complex workflows

Developing isolated execution environments and multi-agent team communication protocols

// getting started

To begin, clone the repository and install the necessary dependencies using 'pip install -r requirements.txt'. Configure your environment by copying the example file to '.env' and adding your ANTHROPIC_API_KEY. You can then explore the progression by running the scripts in the 'agents/' directory, starting with 's01_agent_loop.py' and moving toward the full implementation in 's_full.py'.