learn-claude-code: what it is, what problem it solves & why it's gaining traction

What it solves

This project provides a comprehensive educational framework for building "harnesses" for AI agents. It addresses the misconception that agency is created through prompt-chaining or no-code workflow builders, arguing instead that agency is a property of the model, while the harness is the operational environment (tools, knowledge, and permissions) that allows the model to act.

How it works

The project is structured as a series of 20 progressive lessons that build upon a single, constant "agent loop." Each lesson introduces a new mechanism to enhance the agent's capabilities without changing the core loop. These mechanisms include:

• Tooling: Implementing atomic and composable tools (e.g., bash, file I/O).
• Context Management: Implementing context compaction and memory subsystems to handle long sessions.
• Task Orchestration: Creating task systems with dependency graphs and background execution.
• Collaboration: Setting up agent teams with async mailboxes and shared communication protocols.
• Governance: Establishing permission boundaries and approval workflows.
• Integration: Using the Model Context Protocol (MCP) to plug in external capabilities.

Who it’s for

This is designed for "harness engineers"—developers who want to move beyond simple prompt engineering and learn how to build the professional-grade infrastructure required to deploy autonomous AI agents in specific domains, particularly software engineering.

Highlights

• Progressive Curriculum: A 20-step path from a simple bash-enabled loop to a comprehensive agent harness.
• Harness-Centric Philosophy: Focuses on the operational environment (the "vehicle") rather than trying to engineer intelligence into the model (the "driver").
• Concrete Implementations: Includes runnable code.py files and narrative READMEs for each lesson.
• Architectural Patterns: Teaches patterns for subagent isolation, on-demand skill loading, and worktree isolation.